Skip to content

Hugging Face Blog

feed🧠 AIhuggingface.co β†—

61 posts in the catalog

Each feed is checked for new posts. Every post is read, and those that belong here join the catalog. What Trove read and did not keep stays on file, not shown.

huggingface.coarticle

Granite Switch 4.1.1 β€” Update models like you update software

Granite Switch 4.1.1 is a backwards-compatible patch release of IBM's modular Switch architecture, which configures a language model with specialized, cache-efficient adapters on one endpoint. It updates the Granite Speech ASR model (10 languages, native punctuation and capitalization) and the Decision adapter in granitelib-core, measured on the Jev Decision Index test set. Only the new adapters were retrained; the base model is unchanged. IBM plans tutorials and an Adapter SDK for third-party adapters.

Hugging Face
IBMGranite SwitchGranite Speech+2

models, multimodalΒ· 2026-10-29T13:37:07+00:00

huggingface.coarticle

Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready

Project ProtoStar is an initiative by Carnegie Mellon, Purdue, Georgia Tech, Arizona State, George Washington University, Astera Institute, ORCA and Hugging Face to release publicly funded science datasets, models and code in AI-ready formats. Hugging Face reports over 17,000 public science datasets and tripled sharing in 2026. The first challenge, run by CMU from 2 to 13 November 2026, asks teams to forecast climate indicators for 210 US cities.

Hugging Face
Project ProtoStarCarnegie Mellon UniversityPurdue University+15

ai-policy, training, evalsΒ· 2026-10-29T00:00:00+00:00

huggingface.coarticle

Welcome Le Chonk: Mistral Large 4 on Hugging Face

Mistral Large 4 (Le Chonk) is a multimodal, multilingual Mixture-of-Experts model with about 1T total and 49B active parameters and a 1M-token context, released under the Mistral Ecosystem Open-Weight License. The post covers its MLA and DeepSeek Sparse Attention architecture, a 1.6B Pixtral encoder, and benchmarks including 82% on the Artificial Analysis Cyber Index and 93% on Cybench. It gives run recipes for transformers, vLLM, llama.cpp, TRL fine-tuning and Inference Endpoints.

Mixture of ExpertsTransformersvLLM+11
Mistral Large 4MistralArtificial Analysis+15

models, inference, evalsΒ· 2026-10-29T00:00:00+00:00

huggingface.coarticle

NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt

Nemotron 3.5 Transcribe is an open-weight, promptable speech model that handles transcription in 40 languages, translation, diarization for up to eight speakers, and audio Q&A. It has about 4B active of 30B parameters and tops open-weight ASR and diarization leaderboards. The post also covers the Granary v2 curation pipeline (463,062 hours, 48 languages) and a medical fine-tuning case study that cut held-out WER by 45.8%.

NeMo SpeechNeMo CuratorvLLM+11
NVIDIANemotron 3.5 TranscribeGranary v2+16

models, multimodal, training, evalsΒ· 2026-10-28T14:58:17+00:00

huggingface.coarticle

Build specialized multimodal models with LFM3

LFM3 is Liquid AI's open-weight family of six ~2B-parameter models (text, vision-language, audio, omni and two d1 decision models), all branched from LFM3-1.7B-Base. The base model uses a tri-hybrid backbone of 13 gated short convolutions, 5 sliding-window attention layers and 6 global attention layers, pre-trained on 22T tokens with a 256K context. The post gives transformers snippets, benchmark comparisons against LFM2.5 and Qwen3.5, and a robotics fine-tuning example.

TransformersHugging FacePyTorch+7
Liquid AILFM3LFM3-1.7B+29

models, multimodal, trainingΒ· 2026-10-27T17:16:52+00:00

huggingface.coarticle

LitPivot: Developing research ideas through dynamic contextualization within the literature landscape

LitPivot is a system from Ai2 and the University of Pennsylvania that helps researchers revise early research ideas when reading prior work changes their direction. It splits a proposal into problem, contribution and evaluation facets, groups papers from Asta Paper Finder per facet, and offers critiques with supporting excerpts and suggested rewrites. In a 17-researcher study, domain experts rated LitPivot ideas as better grounded in the literature than ideas made with a text editor and an AI chat tool.

Claude CodeLLM
Ai2University of PennsylvaniaLitPivot+10

research, llm-appsΒ· 2026-10-23T16:15:50+00:00

huggingface.coarticle

Does Your Model Run on AWS Trainium? Look It Up

The torch-neuronx model support Space is a Hugging Face and AWS Neuron lookup that shows whether a Transformers model runs unmodified on AWS Trainium2. Each result lists core count, wall time and peak memory, plus the run.py script and a public NEFF compile cache. Tests generate 32 tokens on up to 16 cores. Results are in the aws-neuron/torch-neuronx-model-support Parquet dataset.

AWS Trainiumtorch-neuronxTorchNeuron+5
AWSQwen3

inference, ai-infraΒ· 2026-10-23T00:00:00+00:00

huggingface.cowriteup

Turning Database Grading into RL Rewards: Post-Training Agents with ThinkingBox

ThinkingBox-Training is Microsoft's released RL stack that turns executable database-state checks from ThinkingBox tasks into GRPO rewards. Fifty GRPO steps on 187 curated tasks raised Qwen3.8-27B pass@1 from 51.70% to 60.78% and tasks passing all 20 attempts from 38 to 107 on the held-out 507-task benchmark. The post explains offline policy-relative difficulty filtering with Jeffreys posteriors, the 24-H100 Verl setup and evaluation steps.

GRPOQwen3.8-27BVerl+9
MicrosoftThinkingBoxThinkingBox-Bench+11

training, agents, evalsΒ· 2026-10-22T22:24:34+00:00

huggingface.coarticle

OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

OlmoEarth v1.3 is an open satellite embedding model from Ai2 that reads a year of Sentinel-1, Sentinel-2 and Landsat data and outputs a 128-dimensional vector per 10 m pixel (the first 64 dimensions also work alone). Its embeddings beat AlphaEarth Foundations and Tessera v2 on 20 of 27 evaluations. Precomputed global embeddings, a browser Embeddings Explorer and pre-training cost (about $2,720 of H100 time) are covered.

TransformerHugging FaceGeoTIFF
OlmoEarthAi2AlphaEarth Foundations+8

models, multimodal, evalsΒ· 2026-10-22T16:11:31+00:00

huggingface.coarticle

NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required

NVIDIA Kumo Relational is a pre-trained relational foundation model that predicts outcomes such as churn or fraud directly on multi-table databases through in-context learning, using up to 10,000 context examples and no task-specific training. NVIDIA reports it beats supervised relational deep learning and XGBoost pipelines across 41 tasks on 15 databases, and beats AutoGluon by 8% MRR on SAP SALT. Code is Apache 2.0, weights are on Hugging Face.

structured-data-modelsNVIDIA NIMPredictive Query Language+6
NVIDIAKumo RelationalKumo Tabular+6

models, research, evalsΒ· 2026-10-21T18:26:30+00:00

huggingface.coarticle

How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications

VocabX is a vocabulary adaptation method from Dharma-AI that selects sub-word and multi-word "macro" tokens by serving cost instead of frequency. On a Qwen3-4B Portuguese essay grader, response length fell from 350 to 101 tokens, the LM head shrank from about 152,000 to 65,000 rows, and measured cost dropped 77.0% on an RTX 5090 with vLLM. The bench score rose from 0.703 to 0.810. It beat FVT, MWT and in-place BPE across 15 model and domain combinations.

Qwen3vLLMLlama+3
Dharma-AIVocabXQwen+1

inference, training, researchΒ· 2026-10-21T17:35:36+00:00

huggingface.cowriteup

Profiling TorchNeuron (Part 1): System-level profiling on AWS Trainium.

Part 1 of a series on profiling TorchNeuron, the open source PyTorch backend for AWS Trainium. It shows how to configure torch.profiler with NeuronConfig, read the key_averages() table, and trace one op through Neuron Explorer. Three RMSNorm variants are compared: naive eager (2.355 ms device time), torch.rms_norm (403 ΞΌs) and torch.compile (375 ΞΌs).

PyTorchTorchNeuronAWS Trainium+8
AmazonHugging Face

inference, ai-infraΒ· 2026-10-21T16:04:43+00:00

huggingface.coarticle

Granite Speech 5.0: Transcribe, Translate, and Attribute more Accurately and Faster

Granite Speech 5.0 is IBM's update to its open speech model collection: the new granite-speech-5.0-2b, a multilingual granite-speech-5.0-470m-turboctc, and a revised granite-speech-4.1-2b-plus. The 2B model ranks #2 among open-weight models on average English WER (5.06%) on the Open ASR Leaderboard. Turbo CTC reaches 19,870 RTFx, and the 4.1 Plus revision cuts long-form WER 24%. All are Apache 2.0.

TransformersvLLMllama.cpp+5
IBMGranite Speechgranite-speech-5.0-2b+5

models, multimodal, evalsΒ· 2026-10-21T14:58:18+00:00

huggingface.coarticle

NVIDIA Kumo Forecast Combines Pretrained Time-Series Representations with Historical Pattern Retrieval and Interpretability

NVIDIA Kumo Forecast (formerly NV-Tesseract Forecasting) is a multivariate time-series forecasting model built on the MOMENT-1-large Transformer backbone (about 340M parameters) with a trainable forecasting head. Its DARR mode blends direct predictions with k-nearest-neighbor retrieval over historical context, and an interpretability framework gives per-horizon lag and feature attributions. Inputs are 256 to 2048 numeric observations; code and weights are Apache 2.0.

TransformerHugging FacePython+1
NVIDIAKumo ForecastNV-Tesseract Forecasting+4

models, researchΒ· 2026-10-21T12:51:24+00:00

huggingface.coarticle

Announcing SimReady Central: An Open Library of Simulation-Ready 3D Assets for Robotics and AI Factories

SimReady Central Library is an NVIDIA-led, free and ungated collection of more than 2,000 simulation-ready OpenUSD assets on Hugging Face, covering robot arms, grippers, warehouse equipment, racks and cooling gear. Assets are validated against SimReady Foundation profiles, and NVIDIA USD Search lets developers and agents find them by natural language. The post shows GPT-6 Astra building a pick-and-place scene in Isaac Sim and comparing AI factory layouts.

OpenUSDNVIDIA USD SearchIsaac Sim+5
NVIDIASimReady CentralSimReady Foundation+14

training, agentsΒ· 2026-10-21T10:59:42+00:00

huggingface.coarticle

NVIDIA Cosmos3-H-Dreams: Bringing Clinical Surgery to Life Through Generative Simulation

Cosmos3-H-Dreams is NVIDIA's real-time generative surgical simulator, trained on clinical cholecystectomy video and kinematics from the CMR Surgical Versius platform. It builds on the 16B Cosmos3-H-Embodied-Nano, distills a bidirectional teacher into four-step and two-step causal students with Self-Forcing, and streams through FlashDreams on one RTX PRO 6000 GPU. Checkpoints and a recipe for custom data are released.

FlashDreamsFadeMemHugging Face+1
NVIDIACosmos3-H-DreamsCosmos-H-Dreams+6

models, multimodal, trainingΒ· 2026-10-20T18:17:47+00:00

huggingface.coarticle

AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms

AnyLanguageModel 1.0 is a Swift package that gives apps Apple's Foundation Models API across the system model, llama.cpp, MLX, Core ML and hosted providers. This release post covers tool calling and guided generation on every provider, image input, and hosted support for OpenAI, Anthropic, Gemini, Ollama and Hugging Face Inference Providers. It also ships swift-huggingface 1.0 and swift-xet 1.0, and lists small open-weight models that fit on iPhone and Mac.

SwiftFoundation ModelsMLX+11
AnyLanguageModelswift-huggingfaceswift-xet+11

inference, llm-apps, mcpΒ· 2026-10-20T00:00:00+00:00

huggingface.coarticle

MPS in PyTorch 2.15

PyTorch 2.15 MPS release post from the Hugging Face PyTorch team, covering the Apple silicon backend. It adds float8_e4m3fn tensors with F.scaled_mm, a native grouped matmul that speeds MoE decoding 5 to 6.5x (Qwen3.6-35B-A3B from 9.4 to 61.3 tokens/s on an M3 Ultra), and an 8x faster softmax. Wan2.2 VAE decode peak memory falls from 270 GiB to 8 GiB. It previews 2.16 attention on Metal.

PyTorchMetalMPSGraph+9
Hugging FaceAppleQwen3.6-35B-A3B+6

inferenceΒ· 2026-10-20T00:00:00+00:00

huggingface.coarticle

Hugging Face ❀️ TorchTPU: TPUs, Natively

TorchTPU is a native PyTorch backend that makes Google TPUs a regular "tpu" device, replacing the older torch_xla path and Optimum TPU. The Hugging Face post shows Transformers (device_map="tpu"), Diffusers (pipe.to("tpu")) and Pallas kernels from the Kernels Hub running on it. It reports up to 21.5x faster decoding on Qwen3-30B-A3B at batch 64 and a 96.5% TPU test pass rate.

PyTorchTorchTPUTransformers+11
GoogleQwen3.5Qwen3+2

inference, ai-infraΒ· 2026-10-20T00:00:00+00:00

huggingface.coarticle

TorchNeuron: running HF models on AWS Trainium is easier than ever

TorchNeuron is an AWS PyTorch plugin that exposes Trainium as a native "neuron" device with eager mode, torch.compile, DTensor and FSDP2, entering public beta on 20 October 2026 (pip install torch-neuronx). The post shows Transformers text generation with tensor parallelism, TRL SFT of Qwen3-8B on 32 NeuronCores (TP 4 x FSDP 8), and Diffusers TP inference. It says 89% of the top 100 text-generation architectures work, versus 46% with Optimum Neuron v0.4.5.

TorchNeuronAWS TrainiumPyTorch+12
AmazonQwen3.8-27BQwen3+1

inference, trainingΒ· 2026-10-19T00:00:00+00:00

huggingface.coarticle

Scaling Indic Pretraining: 2.8 Trillion Tokens and an Open Data Generation Pipeline

Announces Nemotron-pretraining-Indic-v1, about 2.8T tokens across 22 Indian languages (2,601B native-script, 224.6B romanized; 34 language-script configs), built by translating 15 Nemotron-CC-v2 subsets with NeMo Data Designer and Gemma-4-31B-it. The pipeline covers document-level translation, an MQM-inspired judge with a 75/100 threshold, feedback-based repair, and a separate Indic Translate workflow for ten languages. Continued pretraining of Nemotron 3.5 Super VL with 10B Indic tokens raised all five Indic benchmarks by 4.98–9.82 points. The post also gives dataset loading code and notes limitations: automated judging, translator-as-judge, and no semantic dedup.

NeMo Data DesignerGemmaHugging Face+2
NVIDIAAI4BharatIIT Madras+21

training, models, evalsΒ· 2026-10-16T01:44:47+00:00

huggingface.coarticle

Welcome Beam: a 501B open model you can own, customize, and scale

Announces Reflection's Beam-1.0-501B-A23B, an Apache 2.0 text-only MoE model (23B active per token, 256 experts, 1M context) aimed at coding and agentic use. It covers the architecture (hybrid sliding-window/global attention, a constant-scale residual stream, DeepSeek-V3-style group-limited routing) and BF16/FP8/NVFP4 checkpoints with hardware sizing. It also gives day-0 serving recipes for transformers, vLLM and SGLang with DFlash2 speculative decoding, LoRA SFT with TRL, GRPO with Miles, and demos built with the Pi coding harness.

TransformersvLLMSGLang+12
BeamReflectionDeepSeek+15

models, inference, trainingΒ· 2026-10-16T00:00:00+00:00

huggingface.coarticle

Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint

IBM announces Granite-Switch-4.1-8B and -30B. Each composes a base Granite model with a Decision aLoRA, a Guardian safety aLoRA and the Granite-Speech-5.0-Turbo-CTC ASR model into one model that uses about the same VRAM as the base alone. Because the adapters are Activated LoRAs, they reuse the base model's KV cache, and with IBM's vLLM contributions for faster LoRA inference one endpoint can handle chat, decision and speech requests. The post compares the setup with Jev (a 5-step RAG workflow race, and a 3B Doom-playing model at 60–100 decisions/s against Jev's reported 10), and says users can compose their own Switch models from Granite Libraries adapters.

vLLMLoRAActivated LoRA+4
IBMGranite Switch 4.1Granite-Switch-4.1-8B+9

models, inference, multimodalΒ· 2026-10-15T20:34:01+00:00

huggingface.coarticle

Apertus 1.5 Release Notes

Release notes for Apertus 1.5, a fully open (weights, data, training details) continued pretraining of the 8B and 70B Apertus models on 4T and 2T additional tokens. It adds early-fusion image and audio understanding via discrete tokenizers (Emu3.5, WavTokenizer) without architectural change, a thinking mode, 262k context, and better tool use. It documents a four-stage pretraining and five-stage post-training pipeline (SFT, long-context SFT, RLVR, DPO, charter-alignment SDPO), plus evaluation via the MLLM-eval-suite and Every Eval Ever (EEE) Eval Cards, released checkpoints and datasets.

Hugging Facelm-evaluation-harnesslmms-eval+2
ApertusApertus 1.5Emu3.5+17

models, training, multimodal, evalsΒ· 2026-10-15T09:08:45+00:00

huggingface.coarticle

Welcome FLUX.3 Image Klein in Diffusers

Hugging Face blog post introducing Black Forest Labs' open FLUX.3 Image Klein (distilled and base checkpoints) in Diffusers. It covers text-to-image, multi-reference image-to-image and bounding-box layout control, FP8/NVFP4 single-file quantized weights, DreamBooth LoRA training scripts for generation and editing, memory/speed figures (about 2 s per 1024Γ—1024 image on an H100), and a detailed architecture tour: Qwen3-VL text encoder, two-stage per-stream then joint transformer blocks, per-token timestep modulation, 4D RoPE, and flow-matching denoising. It also shows how to customise the Modular Diffusers pipeline.

DiffusersModular DiffusersTransformers+9
FLUX.3 Image KleinBlack Forest LabsFLUX.2+4

models, multimodal, trainingΒ· 2026-10-15T00:00:00+00:00

huggingface.coarticle

Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta

Hugging Face's launch post for Meta's Muse Spark Open, an open-weight 816B-parameter mixture-of-experts model (about 42B active per token). It has 256 experts with 8 routed per token, a 1M-token context, and perception and audio encoders for image, video and audio input. The post gives the architecture details, BF16 and NVFP4 checkpoints, and run recipes for transformers, llama.cpp, SGLang, vLLM, MLX and Inference Endpoints. It also covers LoRA SFT of the full model with TRL on 64 H100s and agentic demos (Pi 3D asset generation, an OpenEnv browser agent, a voice-driven shopping agent).

Mixture of ExpertsTransformersvLLM+10
MuseMetaMuse Glimmer+4

models, inference, multimodalΒ· 2026-10-15T00:00:00+00:00

huggingface.coarticle

HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost

Ai2 announces HiRO-ACEv2, an update to its AI climate emulator. ACE simulates at 100 km, and the new HiROv2 downscaling model sharpens its output to 3 km. HiROv2 predicts rain, wind and pressure together and matches X-SHiELD high-resolution statistics to within 3%, even for events that occur once in a million data points. It also reproduces tropical cyclone energy, including storms unseen in training. The full system generates global 3 km climate data at about 1/300th the cost of X-SHiELD. Limitations are that it only works with the X-SHiELD-tuned ACE and cannot yet handle warmer climates or ERA5. Models, tutorial notebooks and a web demo are available.

HiRO-ACEACEX-SHiELD
Ai2HiRO-ACEv2HiROv1+3

models, researchΒ· 2026-10-14T16:12:42+00:00

huggingface.coarticle

Run fast, private agents on-device with LFM2.5-24B-A2B

Liquid AI releases LFM2.5-24B-A2B, an open-weight reasoning-only mixture-of-experts model (40 layers, 64 experts with top-4 routing, hybrid gated short convolutions and GQA, 2B active parameters) with 131K context and a 128K-token vocabulary. Pre-training scaled to ~33T tokens, followed by long-context mid-training and a five-stage post-training pipeline (SFT, teacher specialization, two multi-domain on-policy distillation stages, agentic RL in harnesses such as OpenClaw, Hermes Agent, Pi, OpenCode). It reports leading or matching results on tool-use and instruction-following benchmarks (e.g. LiveClaw 48.09 vs 38.19 for Qwen3.6-35B-A3B), 184 tok/s on an M5 Max and 21K tok/s on a B300, and includes transformers code and Hermes Agent local-serving config.

Mixture of ExpertsTransformersvLLM+12
Liquid AILFM2.5-24B-A2BLFM2-24B-A2B+18

models, agents, inferenceΒ· 2026-10-13T17:05:48+00:00

huggingface.cowriteup

Impactful scheduling for GPU clusters

Ai2's AI Infrastructure team describes replacing a priority-based scheduler on thousands of H100/B200/B300 GPUs with GPU time budgets, hierarchical fair-share over a 7-day lookback window, and a "scheduling contract" where workloads declare a minimum runtime (capped at 8 hours) to enable time-slicing. Results over 30 days: 98% of owed GPU hours delivered, occupancy steady at 98%, debug-workload p90 queue wait down from about 2 hours to 30 seconds, and 74% fewer repairs needing a human. It also covers a simulator used to predict outcomes before rollout, plus problems that remain: interactive sessions hurt by preemption and possible capacity fragmentation.

SLURMYARNHadoop+3
Ai2NVIDIADominant Resource Fairness+1

ai-infra, trainingΒ· 2026-10-09T15:20:29+00:00

huggingface.coarticle

The model that didn't exist, so you made it yourself

A Hugging Face blog post describing hands-on use of "ML Intern," an agent in HuggingChat that plans a training job, asks for a spending budget, runs a smoke test, then fine-tunes, evaluates and publishes a model to the Hub. The author walks through six concrete projects β€” a citrus-disease VLM fine-tune of Qwen3.5-2B, a Huggy character LoRA on FLUX.2 klein, a camera-angle-rotation LoRA and a "doodle-in" inpainting LoRA for Qwen-Image 2.1, and two distilled text-to-image/text models (a 0.8B prompt-rewriter distilled from a 9B teacher, and a 4-step distilled Agate Preview 002) β€” with datasets built, zero-shot baselines, step counts, hardware (A10G/A100), and total compute costs (roughly $2–$37 per project). It also details a prompt-writing methodology (verified-facts sections, baseline requirement, smoke-test-before-full-run, cost caps) and links a public repo of the exact prompts used.

Hugging FaceLoRAGGUF+5
ML InternQwen-Image 2.1Qwen3.5-2B+12

agents, training, promptingΒ· 2026-10-08T00:00:00+00:00

huggingface.coarticle

Introducing Falcon ASR

TII (Technology Innovation Institute) announces Falcon-ASR, a 1.6B-parameter automatic speech recognition model focused on the Emirati Arabic dialect, which also supports MSA, other Gulf dialects, English, French, Spanish and Portuguese under one set of weights. It reports 20.92% average WER across six Arabic benchmarks on the Open Universal Arabic ASR Leaderboard (beating the best published leaderboard result of 23.17%), 22.73% WER / 10.19% CER on an internal Emirati evaluation (beating Qwen3-Omni by 4.07 points), and 5.74% mean WER on the Hugging Face Open ASR Leaderboard's English test sets. The model builds on the team's Falcon3-Audio work, supports word-level timestamps, and is available via a Hugging Face demo space with API access planned.

Hugging Face
TIIFalcon-ASRFalcon3-Audio+7

models, multimodal, evalsΒ· 2026-10-07T13:21:03+00:00

huggingface.coarticle

One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO

NVIDIA describes how it specialized Nemotron 3 base models into domain experts for competitive programming (IOI) and olympiad mathematics (IMO) using a four-step recipe: strong base model, curated domain data with reasoning traces, SFT/RL post-training, and an inference loop (GenCorrect for IOI, generate-verify-refine for IMO) that searches and refines candidate answers. Nemotron-3-Nano-CC and Nemotron-3-Ultra-CC reached gold-level IOI 2026 scores (535.4/600), while SFT and RL checkpoints of Nemotron 3 Ultra combined to score 30/42 on IMO 2026, exceeding the gold threshold with no formal prover or external tools. The post releases training datasets, checkpoints, the Nemotron-IMO-Bench benchmark, and the NeMo-Skills inference pipelines on Hugging Face.

Hugging FaceMixture of Experts
NemotronNemotron 3Nemotron-3-Nano-CC+11

training, evals, modelsΒ· 2026-10-07T12:45:31+00:00

huggingface.coarticle

Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance

TII describes building Falcon-Emirati-7B, a dialect-specialized LLM fine-tuned from Falcon-H1-Arabic (a hybrid Mamba+Transformer architecture) to handle Emirati Gulf Arabic rather than Modern Standard Arabic, covering its data pipeline (authentic dialect web crawl, MSA cultural text, glossary-constrained synthetic data) and training-stage ablations. It reports 84.83% on the Alyah benchmark (beating larger multilingual and Arabic models), plus LLM-as-judge open-ended generation scores showing it is far more likely than competitors (ALLaM-7B, Jais-2-8B-Chat, Fanar-2-27B-Instruct, gemma-3-27b-it) to actually answer in Emirati dialect rather than defaulting to MSA, and 85.57% accuracy on the ArabCulture-Dialogue UAE cultural benchmark.

TransformerHugging FaceMamba
Falcon-Emirati-7BFalcon-H1-ArabicFalcon-H1+8

models, training, evalsΒ· 2026-10-06T06:44:39+00:00

huggingface.coarticle

The Agent Said It Was Done. The Database Disagreed.

ThinkingBox is an agent sandbox and ThinkingBox-Bench is a 507-task stateful business-workflow benchmark (retail, auto insurance, etc.) that grades agents by checking the terminal backend database state and side effects they leave behind, not just whether tool calls succeeded or the final reply sounds right. Across 121,680 trials over 12-18 LLMs run 20 times each, the post finds most failures are clean-looking: 67% of failed trials still terminated without a tool error, yet executable checks caught wrong field values, missing effects, or unintended extra effects. It also reports pass@1 vs. 20/20 consistency gaps (e.g. Kimi K3 solves the most tasks at least once but is among the least consistent, while Claude Opus 5 and Claude Opus 5.5 are the most dependable) and cost-per-success/cost-per-dependable-task metrics, and ships the benchmark through an OpenEnv interface for reproducible evaluation.

Model Context ProtocolHugging FaceDocker+4
MicrosoftThinkingBoxThinkingBox-Bench+21

evals, agentsΒ· 2026-10-03T22:56:48+00:00

huggingface.coarticle

Open-sourcing AstaBrief, the fast report-generation model in Asta

Ai2 open-sources AstaBrief 8B, a Qwen3-8B-based model fine-tuned to generate cited scientific reports in a single pass (rather than section-by-section), used as the "Fast mode" in their Asta research platform alongside a Claude-powered "Thinking mode." The post details the training recipe (SFT on 47K filtered real-query reports plus DPO on 6K judge-agreed preference pairs), the statistics-based filters (citation density, citation diversity, citation relevance, output/input token ratio) that most improved grounding, and evaluation against SQABench-CS2 and DeepScholarBench, showing AstaBrief is ~3.5x faster than the proprietary pipeline while remaining competitive on citation accuracy and answer quality.

Qwen3Hugging FaceDPO+1
AstaBriefAstaAi2+14

models, training, evalsΒ· 2026-10-02T15:19:50+00:00

huggingface.coarticle

AutoSynthData: Generating Training Data for Enterprise Agents

AutoSynthData is ServiceNow CoreAI's pipeline for generating synthetic training data for enterprise agents: it evaluates a target model against a stronger teacher in a given environment, distills capability gaps into sanitized "specification cards," then generates and validates new tasks (system specification, user prompt, verifier) via a target-then-multiply process with positive/negative verification, critique-repair loops, and batch-level coverage review. On EnterpriseOps Gym, fine-tuning Gemma-4-26B-A4B-it on AutoSynthData-generated samples improved Pass@1 by 7.2 points (35% relative) on the Hybrid domain and from 18.77% to 27.18% on ITSM, closing 59% of the gap to the teacher model on Hybrid.

Hugging FaceGemma
ServiceNowEnterpriseOps GymGemma-4-26B-A4B-it+2

training, evals, agentsΒ· 2026-10-02T04:01:31+00:00

huggingface.coarticle

Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

Allen AI's Olmo-core 3 is an open training framework redesigned to scale mixture-of-experts (MoE) models into the trillion-parameter range while preserving throughput. It replaces the earlier FSDP-based MoE implementation with a distributed-data-parallel system combining expert parallelism, pipeline parallelism and a distributed optimizer, plus optimizations like rowwise expert parallelism, GPU-resident routing, grouped GEMM and MXFP8 low-precision support; the post reports benchmarks up to 1.2T total parameters on 512 NVIDIA B300 GPUs (858 TFLOP/s/GPU) and a 2.38T-parameter test with DeepEP v2, alongside training pitfalls like "token gerrymandering" in load-balancing metrics. It underpins the next-generation Olmo model and is released openly as infrastructure, not just weights.

Mixture of ExpertsMegatron-Core
Olmo-core 3Olmo 3OlmoE+4

training, ai-infraΒ· 2026-10-01T15:01:43+00:00

huggingface.coarticle

Open TTS Leaderboard: Scalable Evaluation for Multilingual Text-to-Speech and Voice Cloning

Hugging Face introduces the Open TTS Leaderboard, which ranks text-to-speech and voice-cloning models using objective, scalable metrics rather than slow human-voting arenas: intelligibility via WER/CER computed with Qwen3 ASR, speed via RTFx (batched) and TTFA (streaming) on H200 GPU/CPU, and speaker similarity via WavLM embedding cosine similarity. It covers multilingual evaluation (Seed TTS Eval, CV3 Eval zero-shot) across English, Chinese, Japanese and Korean, highlights leading open models like Kokoro-82M, supertonic-3, s2-pro, OmniVoice and Fun-CosyVoice3, and includes a Listen tab for human comparison and voting and a Streaming tab benchmarking TTFA, with plans to open-source the evaluation scripts.

Hugging FaceQwen3
Qwen3 ASROpen ASR LeaderboardWavLM+12

evals, multimodalΒ· 2026-09-30T00:00:00+00:00

huggingface.coarticle

NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction

NVIDIA Kumo Tabular is an open foundation model for tabular data (28M–215M parameters, three sizes) that predicts labels for new rows via in-context learning in a single forward pass, with no task-specific training or feature engineering. It uses a Transformer built around column, row, and in-context attention (building on TabICL and TabPFN), Fourier feature cell embeddings, and a length-aware attention temperature to stay accurate on very large tables; it's pretrained entirely on synthetic tables sampled from structural causal models. It tops the TabArena, BeyondArena, TALENT, and ScoringBench leaderboards while running much faster than the next-best model (LimiX-2), and is released under the OpenMDW-1.1 license with weights on Hugging Face and code on GitHub.

TransformerHugging FaceCUDA+1
NVIDIATabArenaBeyondArena+10

models, evalsΒ· 2026-09-29T15:30:38+00:00

huggingface.cowriteup

Getting the Source Right, Not Just the Fact: Source-Aware Verification for MCP Agents

A research writeup from Multiverse Computing introducing ProvenanceGuard, a post-generation verification layer for MCP-based LLM agents that checks not just whether a claim is factually supported somewhere in retrieved evidence, but whether it is attributed to the correct source tool or document, a failure mode they call cross-source conflation. Using local models (MiniLM for source routing, a DeBERTa NLI verifier, and a local LLM for claim decomposition) plus a RARR-style repair loop, it was evaluated on 281 real MCP traces from a medical agent, catching 138 of 139 unsupported claims flagged by human experts and correctly identifying the exact source about 86% of the time on held-out data, outperforming four source-blind baseline checkers. The method has already been adopted in NVIDIA's NVFlow finance-agent grounding stage.

Model Context ProtocolHugging Face
ProvenanceGuardMultiverse ComputingNVIDIA NVFlow+5

evals, research, mcpΒ· 2026-09-29T13:07:00+00:00

huggingface.coarticle

Holo4: powering generalist computer-use agents

H Company's blog post announcing Holo4, a pair of open-weight agentic models (27B dense and 35B-A3B MoE, built on Qwen bases) trained to operate software through GUIs, code execution, MCP, and APIs interchangeably rather than a single interface. It reports benchmark results on OSWorld 2.0 and AutomationBench against frontier models like Claude Opus 5.5 and GPT-5.6 Sol, describes the training pipeline (an internal "Agentic Task Factory" generating ~10,000 verifiable environments/tasks plus SFT+RL), and also releases Holotron4 Nano, the same post-training recipe applied to NVIDIA's Nemotron 3 Nano Omni. Weights (BF16, FP8, NVFP4, GGUF) and full agent trajectories are published on Hugging Face.

Hugging FaceGGUFQwen+2
Holo4Holo4-27BHolo4-35B-A3B+15

models, agents, evalsΒ· 2026-09-28T09:44:05+00:00

huggingface.coarticle

How UK AISI and EvalEval Are Making Benchmark Results Reproducible

UK AI Security Institute (AISI) and the EvalEval Coalition describe a collaboration to make AI benchmark results reproducible: AISI is publishing verified evaluation results, configuration details, and transcript-level data through EvalEval's Evaluation Cards platform, using the shared Every Eval Ever (EEE) reporting schema. The release covers five benchmarks (HealthBench, FrontierMath, Humanity's Last Exam, SWE-Bench Pro, Terminal-Bench 2.0) across six frontier models (Claude Opus 4/4.5/4.6, GPT-5/5.2/5.4), plus two cyber evaluations (Cyber CTFs, The Last Ones), accompanying AISI's paper on how inference compute and evaluation protocol shape benchmark performance.

Hugging Face
UK AI Security InstituteEvalEval CoalitionEvery Eval Ever+16

evalsΒ· 2026-09-22T00:00:00+00:00

huggingface.coarticle

Transformers now runs llama.cpp quants

Hugging Face Transformers can now load GGUF quantized checkpoints directly via from_pretrained(gguf_file=...) and run them at near-llama.cpp speed on Apple Silicon by reusing ggml's Metal kernels through the kernels library, plus generation-loop optimizations such as dropping unnecessary attention masks and deferring stopping checks. The post benchmarks Qwen3.5 dense and MoE GGUF checkpoints against llama.cpp's llama-bench on an M2 Max, shows how to serve GGUF models with transformers serve as an OpenAI-compatible endpoint usable by clients like Jan or Pi, and covers dequantizing GGUF weights for fine-tuning via GgufConfig(dequantize=True). Current limitations: MPS-only packed inference, no batching support yet, and architecture coverage limited to Qwen3.5/Qwen3.8.

Hugging Face Transformersllama.cppGGUF+4
Hugging FaceQwen3.5Unsloth+12

inference, modelsΒ· 2026-09-22T00:00:00+00:00

huggingface.coarticle

tokenizers v1: encode, decode and scaling, measured

Hugging Face details the engineering behind tokenizers v1, a release candidate that claims 3-30x faster single-threaded encoding than v0.23 while producing identical token IDs. Key changes include a hand-written SIMD/bitstream splitter (bitcannon) replacing regex-based pre-tokenization for common BPE patterns (GPT-2, cl100k, o200k, Tekken, DeepSeek), a thread-local word cache for repeated pre-tokens, and an allocation-free BPE merge loop using flat arrays and packed 64-bit comparisons. Benchmarks were run via the open tokbench repository across ten model families, comparing against alternatives like tiktoken and other fast tokenizer libraries.

RustPythonNode.js+1
tiktokenIBMNVIDIA+15

inferenceΒ· 2026-09-21T00:00:00+00:00

huggingface.coarticle

Your Agent Aced the Task. Will It Do It Again?

IBM Research's ALTK-Evolve blog post introduces "consistency guidelines," built on a new diagnostic called the Consistency Analyzer, which resamples an agent's own recorded trajectory to find flip-prone decision points without needing ground truth or a live replay. On AppWorld with a GPT-4.1 ReAct agent, they show a 24.4-point gap between Mean@5 (77.4%) and Pass^5 (53.0% β€” success on all 5 repeated runs), and demonstrate that injecting the generated guidelines halves that gap to 12.0pp while preserving or improving average accuracy, with gains that generalize to similar tasks and to a weaker model (gpt-oss-120b). The post argues Pass^k should be reported alongside Mean@k as a distinct, orthogonal measure of agent reliability, and links an open-source repo and arXiv technical report.

ALTK-EvolveHugging Facegpt-oss
IBM ResearchAppWorldGPT-4.1+2

evals, agentsΒ· 2026-09-15T16:00:44+00:00

huggingface.cowriteup

Async GRPO with LoRA across HF Jobs: a bucket, a proxy, and no NCCL

A Hugging Face engineering blog post describing how TRL's AsyncGRPOTrainer now syncs only a LoRA adapter (not full weights) to vLLM, letting trainer and inference replicas run as separate Hugging Face Jobs connected via a Storage Bucket instead of NCCL. It details a custom proxy that adds auth headers, routes rollouts to the vLLM replica most likely to hold the matching KV-cache prefix, and broadcasts adapter loads atomically across replicas. Five iterative runs (token-budget batching, disabling gradient checkpointing, raising in-flight caps) cut a 500-step Qwen2.5-Math-1.5B GRPO run from 3h27m to 53 minutes.

Hugging FacevLLMLoRA+4
TRLAsyncGRPOTrainerHugging Face Jobs+10

training, inference, ai-infraΒ· 2026-09-10T00:00:00+00:00

huggingface.cowriteup

Rebuilding AUTOMATIC1111 with Gradio Workflow

A Hugging Face blog post walking through Workflow1111, a 73-node graph built with Gradio's gr.Workflow that reimplements most of AUTOMATIC1111's stable-diffusion-webui feature set as a single visual canvas: text-to-image, hi-res fix, image-to-image, prompt matrix, VLM interrogation, detection-to-inpaint masking, ControlNet-style annotators, background removal, PNG metadata round-tripping, and image-to-video, wired from models like FLUX.1-Kontext, Qwen3-4B, Qwen2.5-VL, DETR, AuraSR, BRIA RMBG-2.0, and Wan 2.2 I2V A14B. It contrasts gr.Workflow with ComfyUI, shows that every output node auto-becomes both a REST endpoint and an MCP tool (usable by Claude Code or Cursor), and demonstrates running nodes locally on your own GPU via ZeroGPU (FastVideo/fastvideo-fasth3-preview) instead of hosted inference.

GradioHugging FacePython+5
AUTOMATIC1111gr.WorkflowWorkflow1111+14

llm-apps, multimodalΒ· 2026-09-10T00:00:00+00:00

huggingface.coarticle

NeoMME: an efficient Multimodal-native and Multilingual Encoder

NeoMME is a family of 260M/800M multilingual, multimodal-native encoders from H Company that use a single bidirectional Transformer (not a separate vision tower + causal LM) to jointly process text tokens and raw 32Γ—32 image patches, trained from scratch with a masked discrete-diffusion objective over ~524B tokens. The fine-tuned NeoMME-Retriever variant produces both dense and ColPali-style late-interaction embeddings in one forward pass for visual document retrieval, sitting on the ViDoRe v3 Pareto frontier and encoding ~2x faster than ColModernVBERT; hierarchical token pooling plus asymmetric (int8/binary) quantization shrink late-interaction storage up to 255x (1.5MB to 6kB/page) with under 5% nDCG@10 loss. Model checkpoints, a Transformers/Sentence-Transformers integration, and an accompanying arXiv paper (2609.01657) are released under Apache 2.0.

Hugging FaceTransformervision-language model+1
NeoMMENeoMME-RetrieverH Company+23

models, multimodalΒ· 2026-09-03T13:13:48+00:00

huggingface.coarticle

Training a coding model to paint watercolours with TRL and OpenEnv

A Hugging Face engineering blog post reproducing Surya Narreddi's viral "AI paints watercolours" project in the open: a coding LLM (Qwen3.5-35B-A3B, LoRA) is trained with GRPO via TRL to write p5.brush JavaScript sketches that render as watercolour flower paintings, using an OpenEnv RL environment with headless-Chromium rendering, an HPSv3 preference-model reward and a Qwen3-VL-30B-A3B-Instruct pairwise judge scored against a hand-curated 178-image reference pool. It documents the full pipeline (HF Jobs, Spaces, Inference Providers), the LoRA fix needed for MoE models (all-linear targeting), infra bugs found (an OpenEnv websocket caching bug fixed upstream, reward-noise from silent render failures), three reward-mix ablations, and open-sources the dataset, environment, training scripts, rollouts and trained models.

TRLOpenEnvHugging Face+8
Surya NarreddiHPSv3Qwen3.5-35B-A3B+12

training, multimodalΒ· 2026-09-03T00:00:00+00:00

huggingface.cowriteup

Fine-tuning a 350M Model for Better Structured Outputs in 100 GRPO Steps

A Hugging Face blog walkthrough showing how to GRPO-fine-tune LiquidAI's LFM2.5-350M model with TRL and a LoRA adapter (targeting LFM's hybrid attention/convolution modules) using three custom reward functions β€” JSON/YAML format compliance, field-count matching, and JSON Schema validation β€” trained on ~500 samples from nvidia's Nemotron-RL-instruction_following-structured_outputs dataset for just 100 steps. On the IFStruct benchmark, the fine-tuned model's overall pass rate rises from 22.6% to 29.7% (JSON specifically from 18.0% to 31.9%), narrowing the gap to the much larger Qwen3.5-2B (33.15%). The post also covers converting the merged model to GGUF and serving/evaluating it locally via llama.cpp's OpenAI-compatible server.

Hugging FaceGRPOLoRA+4
LFM2.5-350MIFStructLiquid4All/ifstruct+4

training, evalsΒ· 2026-09-03T00:00:00+00:00

huggingface.coarticle

Give Your Coding Agents a Memory You Own

funes is an open-source, single-binary memory layer that turns the session traces of coding agents (Claude Code, Codex, Pi, Hermes) into a durable, queryable memory instead of an unsearchable archive. It runs locally by default β€” chunking, embedding and reranking turns with a pinned local model into a Lance dataset, combining vector and BM25 search with cross-encoder reranking and recency weighting β€” and can optionally sync to a private-by-default Hugging Face dataset so a memory follows a user across machines or is shared across a team, with credential redaction before publishing. The post also reports a "handoff-vs-recall" benchmark showing recall from raw traces is cheaper and more reliable than context compaction or written handoffs for resuming long investigations.

Claude CodeCodexHugging Face+2
funesPiHermes+2

coding-agents, ai-infraΒ· 2026-09-03T00:00:00+00:00

huggingface.coarticle

Introducing @huggingface/kernels: 200+ WebGPU Kernels for Local AI

Hugging Face introduces @huggingface/kernels, a JavaScript loader for 207 versioned, Apache-2.0 licensed WebGPU kernels (matmuls, normalizations, attention primitives, quantization ops, etc.) published on the Hub, each with a manifest, correctness tests, benchmarks, and parameterized WGSL shader templates. They also launch Fleet, a browser-based benchmarking/testing tool that crowdsources correctness and performance evidence across real-world GPUs. Benchmarked against ONNX Runtime Web on an Apple M4 GPU, the kernels were 2.57x faster by geometric mean (1.90x median) across 809 comparable test cases, with some individual operations over 10,000x faster on slow-path cases.

Hugging FaceWebGPUWGSL+2
Apple M4

inferenceΒ· 2026-09-01T00:00:00+00:00

huggingface.coarticle

The Open ASR Leaderboard Adds Its First Global South Language

Hugging Face's Open ASR Leaderboard adds Monsoon en-IN and hi-IN, its first Indic-language and Global South evaluation sets, built from 4,888 speaker-disjoint contributors across hundreds of Indian districts with 12 recorded demographic and device attributes per speaker. The post shows that models indistinguishable on aggregate WER (within 0.18 points) diverge by up to 1.68 points once results are broken down by region, and introduces OIWER, a lattice-based metric from AI4Bharat that scores Hindi transcripts against multiple accepted spellings instead of a single reference, changing model rankings versus flattened WER. It also details the collection pipeline and open-sources the scoring implementation voi-oiwer.

Hugging Face
Open ASR LeaderboardMonsoonVoice Arena+9

evals, multimodalΒ· 2026-08-28T00:00:00+00:00

huggingface.coarticle

Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers

A detailed Hugging Face blogpost on training and finetuning multi-vector (ColBERT-style late-interaction) embedding models using the new MultiVectorEncoder class in Sentence Transformers. It walks through model selection (finding that "-unsupervised" pre-supervised checkpoints adapt better to new domains than finished ones), dataset formatting, loss functions (CachedMultiVectorMultipleNegativesRankingLoss, MultiVectorDistillKLDivLoss), training arguments, and evaluators, then presents a full worked example finetuning mLateOn-medical on 1M MIRIAD medical question-passage pairs in 14.5 hours on a single RTX 3090, beating over 50 dense/sparse/lexical/multi-vector baselines by +0.062 NDCG@10. It closes with index optimization via HierarchicalTokenPooling and PLAID-style quantization (with Omar Khattab) showing a 13x storage reduction with minimal quality loss.

Sentence TransformersHugging FacePython+5
MultiVectorEncoderColBERTLightOn+20

training, llm-appsΒ· 2026-08-26T00:00:00+00:00

huggingface.coarticle

Granite 4.2 LLMs: How They're Built

IBM's technical write-up on Granite 4.2, its first family of dense decoder-only reasoning LLMs (3B, 8B, 30B), post-trained from Granite-4.1 base models pretrained on ~15T tokens with a five-phase schedule extending context to 512K tokens. It details the SFT data mixture (agentic vs. non-agentic, quality filtering with LLM judges GPT-OSS-120B and Gemma 4), a staged asynchronous-GRPO RL pipeline (RLVR, skill boosters, then agentic SWE/Terminal/Search stages for 8B/30B, finishing with RLHF), and the NeMo-RL/NeMo-Gym infrastructure used to run tool-using RL rollouts at scale, plus FP8/NVFP4/MXFP4/GGUF quantized releases and integration guides for vLLM, SGLang, OpenCode, Pi, and OpenHands.

vLLMSGLangllama.cpp+7
Granite 4.2IBMGranite-4.1+21

models, trainingΒ· 2026-08-25T15:14:14+00:00

huggingface.coarticle

Quantization-Aware Healing: a compressed, 4-bit model that outperforms its full-precision original

Multiverse Computing introduces Quantization-Aware Healing (QAH), a recovery technique for models that have been both structurally pruned and quantized to low bit-width (MXFP4), applied here to a GPT-OSS 120B model compressed to 60B parameters. Unlike quantization-aware training (QAT) or quantization-aware distillation (QAD), which anchor recovery to the degraded recovered checkpoint, QAH distills directly from the original full-precision, full-size teacher via KL-divergence on logits, producing a 4-bit student that beats its own bfloat16 source on 7 of 9 benchmarks and even surpasses the full-size teacher on LiveCodeBench, while reaching peak accuracy roughly 7x faster than QAT and remaining stable rather than degrading with continued training.

Hugging FaceGPT-OSSMXFP4
Multiverse Computinggpt-oss-120bMMLU-Pro+5

training, inferenceΒ· 2026-08-25T11:39:24+00:00

huggingface.coarticle

Wire It, Run It, Deploy It: AI Workflows in Gradio

This Hugging Face blog post introduces gr.Workflow, a Gradio primitive that lets developers describe an AI pipeline as a graph of typed nodes (references for inputs, operators for steps, subjects for outputs) and get a runnable drag-and-drop canvas, an auto-generated REST API per output, and one-command deployment to Hugging Face Spaces from the same definition. Operator nodes can call Hugging Face Inference Providers, other Gradio Spaces, Hub dataset rows, or plain Python functions decorated with @spaces.GPU for in-Space GPU inference via ZeroGPU. Worked examples include an image-editing pipeline using Qwen-Image-Edit, a multi-model media studio chaining FLUX, background removal, and TTS, a parallel fan-out image-generation demo, a Hugging Face dataset profiler using the Datasets Server API, and a GPU-based image animator using Lightricks/LTX-Video via Diffusers.

GradioHugging FacePython+3
Hugging Face SpacesHugging Face Inference ProvidersQwen-Image-Edit+5

llm-apps, ai-infraΒ· 2026-08-25T00:00:00+00:00

huggingface.coarticle

How Hugging Face Inference Endpoints, Jobs, and Buckets Power Search on Papers with Code

A Hugging Face engineering blog post detailing how Papers with Code built a hybrid search system combining PostgreSQL full-text search with pgvector-based dense retrieval, fused via reciprocal rank fusion (RRF). It walks through using Hugging Face Jobs for batch embedding of 110,000+ papers with Qwen3-Embedding-0.6B (using Matryoshka Representation Learning to shrink vectors to 256 dimensions), Storage Buckets as an immutable, checksummed handoff layer between database, compute, and production, and an authenticated Inference Endpoint (backed by Text Embeddings Inference) for low-latency query embedding with graceful fallback to lexical search on cold starts or failures.

Hugging FacePostgreSQLpgvector+5
Papers with CodeQwen3-Embedding-0.6BMTEB+3

ai-infra, llm-apps, inferenceΒ· 2026-08-21T00:00:00+00:00

huggingface.coarticle

Measuring benchmark optimization in speech recognition

Artificial Analysis introduces three diagnostic tests to detect "benchmaxxing" in speech recognition: consensus disagreement (checking if models reproduce known-erroneous VoxPopuli reference transcripts instead of the audio), masked entity retrieval (silencing numbers in audio to see if models hallucinate the expected benchmark answer), and orthographic switching (checking if models match a benchmark's spelling convention, e.g. "Mr." vs "Mister", purely from acoustic cues). Testing 11 open-source ASR models against VoxPopuli and LibriSpeech, they found top-scoring models reproduce reference errors 18-30% of the time and can identify which benchmark they're being tested on from acoustic context alone, with the effect weakening on freshly collected post-training-cutoff audio. The findings led to a new "Benchmark fitting" tab on the Open ASR Leaderboard and open-sourced evaluation scripts.

Hugging Face
Artificial AnalysisVoxPopuliLibriSpeech+6

evals, multimodalΒ· 2026-08-21T00:00:00+00:00

huggingface.coarticle

Up to 3.2x Faster Inference with LFM2.5-DSpark

Liquid AI describes DSpark, a speculative-decoding method combining a DFlash-style parallel draft backbone, a lightweight sequential (Markov-chain) head, and a confidence-scheduled verifier that prunes low-confidence draft suffixes. They release ~300M-parameter DSpark draft models for LFM2.5-2.6B, LFM2.5-1.2B-Instruct and LFM2.5-8B-A1B (Safetensors and GGUF), reporting up to 3.18x GPU throughput (SGLang on H100) and up to 2.87x on-device throughput (llama.cpp/Metal on M4 Max), plus a 57% average cut in function-calling latency, with day-one upstream support in llama.cpp and SGLang.

SGLangllama.cppHugging Face+1
LiquidAILFM2.5-2.6BLFM2.5-1.2B-Instruct+4

inference, modelsΒ· 2026-08-20T16:52:57+00:00