30 results for “Qwen”

Model card for Qwen3.8-27B, a 27B dense vision-language model from the Qwen team with a hybrid Gated DeltaNet / Gated Attention layout (64 layers, 16×(3 DeltaNet + 1 attention)), MTP training, and 262k native context extensible to 1M via YaRN. It targets coding, research and long-horizon agentic tasks, with thinking on by default, reasoning_effort levels (xhigh/medium/low) and preserve_thinking. The card gives deployment recipes for Transformers, vLLM, SGLang and TokenSpeed, recommended sampling parameters, video-token settings for hour-scale video, and benchmark notes (SWE-bench Pro, NL2Repo-Bench, DeepSWE, HLE, MathVision, WebArena-Verified, etc.).

huggingface.corepoAdded 9 Oct 2026🧠 AI

Kolibri-1 is Aleph Alpha's 78B-parameter mixture-of-experts reasoning model focused specifically on German and English, released under Apache 2.0 with weights, a vLLM plugin, and an extensive model card. It supports explicit reasoning-effort levels, Hermes-style tool calling, and a native 262144-token context validated to over 1M via sliding-window attention; the card details its UniBPE German-aware tokenizer, pre-training mid-training and long-context training recipe, MergeMix-based data-mix search, Merlin-Arthur hallucination-mitigation RL protocol, and benchmark results against Qwen3.5, Gemma 4, Nemotron 3 Nano, and Apertus on MMLU-Pro, GPQA, AIME, SWE-bench Verified, Terminal-Bench 2.1, and HLE.

huggingface.corepoAdded 8 Oct 2026🧠 AI

EvoVLM searches over prompts, image preprocessing, generation budgets and answer extraction for frozen Qwen3-VL-2B-Instruct and Qwen3-VL-2B-Thinking, without fine-tuning weights. A local Qwen model ranks a catalog of typed mutations, a trusted renderer produces Python programs, and Pareto selection on accuracy and latency drives two small generations on scientific chart understanding (CharXiv subset). Recorded results on Dev128: Evolved Instruct keeps accuracy with 8.9% lower median latency, and Evolved Thinking gains 16.41 points of accuracy with 10.6% lower median latency. CPU-only replay and fail-closed evidence verification are included. The authors state the limits: small sample, overlap with the evolution set, and results tied to the recorded hardware.

github.comrepoAdded 16 Oct 2026, published 2 Oct 2026🧠 AI

Ternary-Bonsai-2-27B is a 27B-parameter reasoning model (built on Qwen3.8-27B) quantized to true ternary weights ({-1,0,+1}) at ~1.72 bits/weight, shrinking it from ~54GB FP16 to ~5.95GB while retaining 98.2% of FP16 benchmark quality (84.78 vs the conventional sub-4-bit IQ2_XXS build's 72.59). It uses a Hadamard-rotated weight basis and ships two GGUF packing formats (PTQ1_0 dense trits, PQ2_0 2-bit slots) that require a custom llama.cpp/MLX kernel fork since stock llama.cpp cannot decode them correctly. The card includes per-category benchmark breakdowns (math, coding, agentic tool calling), cross-platform throughput/power measurements (Apple M5 Pro/Max, RTX 5090, L4, H100/A100/Blackwell), and an "intelligence density" metric comparing capability retained per GB against FP16 and other quantizations.

huggingface.corepoAdded 29 Sep 2026🧠 AI

Hugging Face collection for Xiaomi's MiMo-V2.6 model family, comprising five open-weight checkpoints: MiMo-V2.6-Pro-RL (1T params), MiMo-V2.6-Flash-RL (311B params), a MiMo-V2.6-Distill-Qwen-9B distillation (vision-language capable), and Pro/Flash MOPD variants. The title claims MIT licensing and a 1M-token context window, positioning this as a large open-weight release spanning full-scale RL-trained models down to a 9B distilled variant usable on consumer hardware.

huggingface.corepoAdded 29 Sep 2026🧠 AI

A curated GitHub list of open-weight "uncensored" and security-fine-tuned LLMs intended for authorized red teaming, pentesting, and security research, covering models like WhiteRabbitNeo's DeepHat V2, BugTraceAI-CORE, CYBER-FROST, and several abliterated Qwen/GLM/DeepSeek variants. It explains the underlying techniques (abliteration per Arditi et al. 2024, QLoRA/DPO fine-tuning, GGUF/FP8 quantization) and links to Hugging Face model cards, benchmark suites (CyberBench, TrustedSec's LLM Attack Benchmark vs OWASP Juice Shop), and deployment platforms for running these models locally or via API.

github.comrepoAdded 2 Oct 2026, published 29 Sep 2026🛡️ Security

A GitHub README curating open-weight "uncensored" and security-fine-tuned LLMs (e.g. WhiteRabbitNeo's DeepHat V2, BugTraceAI-CORE, CYBER-FROST, CyberPal2.0, Lily-Cybersecurity-7B, Dolphin-Llama3-8B, Wizard-Vicuna-13B-Uncensored, huihui-ai's abliterated Qwen variants) with HuggingFace download links, quantization formats (GGUF, FP8, Q4_K_M, Q6_K), and training techniques used (SFT, QLoRA, DPO, abliteration, MoE/MTP architectures). It also includes a glossary explaining abliteration (orthogonalizing the refusal direction out of the residual stream, per Arditi et al. 2024) and references benchmarks like CyberBench and TrustedSec's LLM Attack Benchmark, making it a practical reference for anyone sourcing or deploying refusal-removed models.

github.comrepoAdded 2 Oct 2026, published 29 Sep 2026🧠 AI

Release notes for vLLM v0.30.0, a major update to the open-source LLM inference/serving engine spanning 762 commits. Highlights include new model support (DeepSeek-V4.1-Flash, GLM-5.3-Flash, Kimi K3, Qwen3.8-Flash-Next, K2-Horizon, Cohere Compass and others), a "Fast Start" GPU weight-cache daemon for near-instant engine restarts via CUDA IPC, Gumbel-max watermarking for generated text, a HiSparse host-memory tier for sparse-MLA KV offload, Model Runner V2 improvements (dual-batch overlap, speculative decoding under pipeline parallelism), expanded quantization support (NVFP4, W4A16 DSA, AutoRound), and hardware-specific tuning across NVIDIA, AMD ROCm, Intel XPU and CPU backends.

github.comrepoAdded 28 Sep 2026🧠 AI

OneJev is a multimodal "System One" decision model fine-tuned from Qwen3.5 and Qwen3.8-27B, released in 0.8B, 4B, 9B and 27B sizes. It returns calibrated probabilities for typed questions (yes/no, choice, score) about screenshots, photos, video and text in one forward pass. The repo covers serving with Flide or llama.cpp GGUF, Docker, the 94,707-example OneJev-Data dataset, and SFT configs using cross-entropy plus Brier loss.

github.comrepoAdded 2 Nov 2026, published 27 Sep 2026🧠 AI

Qwen-Planner-Agent is Alibaba's Tongyi MAI Team framework pairing a trained Planner Model with a stateful Harness to execute mobile tasks end-to-end, coordinating tools, persistent memory, reusable Skills, and sub-agents. It introduces Competence-Adaptive Reward-and-Advantage Engineering (CARE), a hybrid supervised+RL training approach, and reports the 27B model reaching 77.05% Overall on MobilePA-Bench at an estimated $2.41 per 1,000 tasks in output-token cost. This repo hosts the technical report/blog and recorded interactive examples rather than model weights or code.

github.comrepoAdded 26 Sep 2026, published 18 Sep 2026🧠 AI

A community-maintained GitHub list cataloging "decision models" (also called System One models): small/fast models that take a state plus caller-declared questions (Noul/true-false, Choice/multi-option, Score/rubric) and return a probability per answer in one forward pass, rather than generating text. It tracks the originating TypeSafe AI Jev model and its /v1/systemone API alongside competing hosted models (Upstage Solar Decide, Liquid AI's d1, OpenAI's Decisions API, Inception's Mercury Decide, Respan's Span-01), dozens of open-weight fine-tunes (Cloudflare's Clef, Perplexity's pplx-decider, AWS Strands Decider, Together AI's Tev1, and many Qwen3.5-based community models), zero-training logit-reading approaches, local runtimes (Ollama, Ollaya, llama.cpp-based servers), SDKs across a dozen languages, and a very long list of agent/coding-tool/MCP integrations (browser agents, code review, routing, game-playing demos) that use these models for fast, typed, in-the-loop decisions. The page is truncated at 40k of ~51k characters, so later sections (further toys, demos, and possibly benchmarks) are not reflected here.

github.comrepoAdded 2 Oct 2026, published 17 Sep 2026🧠 AI

A GitHub repository archiving 19 jailbreak, prompt-injection, persona-conditioning, and response-format-manipulation prompts collected against DeepSeek, Gemini, GLM, Grok, Kimi, Qwen, Sonnet, and ChatGPT. Prompts are kept close to their original form for provenance/reproducibility, positioned as a corpus for red-teaming and robustness evaluation rather than a guaranteed exploit list, with guidance to test in isolated environments and log model version/config/date/response per trial.

github.comrepoAdded 22 Sep 2026, published 17 Sep 2026🛡️ Security

Decider is an open reproduction of TypeSafe AI's System One (Jev) model class: instead of generating text, it reads a state plus a set of typed questions (Choice, Score, or Noul/yes-probability) and returns a calibrated probability distribution for each in one forward pass, with no decoding or parsing. The repo ships 2B/4B/12B and a 35B mixture-of-experts checkpoint (built on Qwen3.5 and Gemma-4 base models), trained on roughly 95 public datasets plus teacher-labelled data, with per-type temperature calibration, GGUF/llama.cpp and vLLM serving paths, an MPS/Apple Silicon port, and a community Snapdragon X Elite NPU port. It benchmarks against TypeSafe's Jev on JevBench and the Decision Index, documenting where it closes the gap (language, tools, retrieval) and where it lags (knowledge-heavy MMLU/GSM8K/GPQA tasks and calibration on hard items).

github.comrepoAdded 2 Oct 2026, published 16 Sep 2026🧠 AI

Index-Translate is Bilibili's Apache-2.0 family of multilingual, multimodal translation models built on Qwen3.5 (2B, 9B and 35B-A3B preview), covering 150 languages with instruction following (glossaries, format preservation, tone). Variants add native image translation with no OCR (vision encoder), Index-Echo for speech subtitles and dubbing, Index-Homura for syllable-controlled translation and Index-NativeLong for full-document translation. The repo provides GGUF, FP8 and NVFP4 quantized builds, vLLM serving scripts, a free OpenAI-compatible API, four new benchmarks (instTrans, MEME, SandGlass, NAtIveLong), and links to three papers including HOMURA and RIVAL (RL for translation).

github.comrepoAdded 10 Oct 2026, published 15 Sep 2026🧠 AI

RepoGuardBench is a benchmark and accompanying paper that measures whether local open-weight coding agents (Qwen2.5-Coder 1.5B-14B via Ollama) stay on the bug-fix task when repository artifacts — README files, issue text, code comments, test logs, and agent-rule files — carry inert prompt-injection instructions. It evaluates five lightweight context- and action-level defenses (D0-D5, including a path-only action gate) across seven metrics in a sandboxed task workspace, finding no statistically clear utility winner among them. The project ships a deterministic no-GPU smoke test, paper-claim verification scripts, and reproducible tables/figures under MIT (code) and CC BY 4.0 (data) licenses.

github.comrepoAdded 22 Sep 2026, published 13 Sep 2026🛡️ Security

Pooled runs one open model (Qwen 3.6 35B MoE, a 27B dense hybrid, smaller Qwen3/SmolLM) across devices in a browser "room". Each tab holds a slice of layers and runs them on its own GPU with custom WGSL kernels. Hidden states pass over WebRTC, and speculative decoding is checked against plain decoding by golden tests. It also ships a Code-mode agent loop with a sandboxed preview, and an npx CLI bridge (@pooled/cli) that exposes the room as an OpenAI Chat Completions/Responses and Anthropic Messages endpoint with grammar-constrained tool calling, so Codex CLI, Claude Code and opencode can use it. The README includes latency-injected benchmarks, a comparison with exo, Petals, llama.cpp RPC and WebLLM, and a threat model.

github.comrepoAdded 16 Oct 2026, published 1 Sep 2026🧠 AI

Reef is an open-source infrastructure stack for continual self-improving agents, wiring together agent inference, scored feedback, training/updates, and versioned delivery into one loop. It supports two learning surfaces: fine-tuning model weights via RL training steps (using Slime and SGLang on top of models like Qwen2.5-1.5B-Instruct) or evolving an agent's harness — prompts, rules, and skills — through its built-in "Reefine" recipe, which lets a model propose and install harness changes from plain-language requests. It exposes an OpenAI/Anthropic-compatible inference endpoint plus a reporting endpoint for scoring interactions, and ships a recipe catalog (e.g. the SAO recipe referencing arXiv:2607.07508) for pluggable learning strategies.

github.comrepoAdded 2 Oct 2026, published 31 Aug 2026🧠 AI

LoongSage is Baidu's open-sourced, production-grade reinforcement learning framework for post-training LLMs, built on Ray for scheduling, Megatron-Core/Megatron-Bridge for training, and SGLang for inference. It targets agentic RL with a zero-intrusion AgentFlow module (automatic multi-turn context, loss masking, trajectory collection), multi-paradigm online distillation (PG-Style, GKD-Style, TopK/full-vocab/JSD KL), fully asynchronous training with partial rollout, and system-level training-inference consistency guarantees (Router Replay, TITO, M2PO, OPSM) that keep probability bias at 1e-4 in MoE/low-precision settings. Recipes are validated on Baidu's internal production runs of DeepSeek-V4-Flash and Qwen3-30B-A3B for math reasoning, agentic code repair, and distillation.

github.comrepoAdded 30 Sep 2026, published 28 Aug 2026🧠 AI

halogen-flash-server is an inference engine that runs Qwen3.8-Flash-Next on AMD Strix Halo (Ryzen AI Max+ 395), serving OpenAI and Anthropic APIs on one port. Version 0.18.0 reports up to 94 tok/s decode and about 1,700 tok/s prefill, with a 1.6 s first token on agent turns. Output at temperature 0 is byte-identical to greedy decode under speculation. It also loads llama.cpp GGUF files and runs small NPU models.

github.comrepoAdded 2 Nov 2026, published 26 Aug 2026🧠 AI

Cumora is a cross-platform team-chat application where AI agents sit alongside humans as first-class participants — holding persistent personas and memory, claiming Kanban work, coordinating with each other, and sending real email — rather than only responding when addressed. Agents can run on Cumora's cloud (per-agent Kubernetes pods driving a multi-hop tool-calling loop over the OpenAI Responses API) or in "Bring Your Own Agent" mode, pairing a local Mac/VPS daemon with an existing coding-agent backend (Claude Code, Codex, Cursor Agent, Gemini CLI, OpenCode, Qwen Code, Google Antigravity, etc.) while keeping provider keys off the server. The repo documents a concrete coordination design — a seen-cursor freshness gate that holds stale replies, atomic claims on units of work, a small-brain triage gate, and a transactional Postgres outbox for realtime fan-out — making it a useful reference for building multi-agent systems that avoid agents colliding on shared work.

github.comrepoAdded 30 Sep 2026, published 17 Aug 2026🧠 AI

shuohao-skills is a set of five chained Claude Code / Codex agent skills that turn a novel into short-drama production assets: novel-outline (adaptation outline), novel-characters (cast bible), novel-art (scene/prop art bible), novel-script (episode scripts with a duration gauge), and novel-storyboard (shot-by-shot storyboards with actual generated shot images and prompts), plus a standalone character-refs skill that generates character reference images (anchor via GPT, rest via Qwen). Each skill is a self-contained directory with a required SKILL.md and a dependency-free scripts/selftest.mjs, installable via a symlink script that auto-detects Claude Code vs codex, and there is a separate report.mjs assembler that merges the five bilingual HTML reports into one navigable page without touching each skill's own render code (handling CSS/JS scoping and image-path rewriting).

github.comrepoAdded 30 Sep 2026, published 6 Aug 2026🧠 AI

This GitHub repository collects verbatim leaked/extracted system prompts — the hidden instructions and guardrails — from a wide range of AI products including Claude.ai, Claude Code, ChatGPT, Codex, Gemini, Grok, Perplexity, Microsoft Copilot, Cursor, Meta AI, Mistral, Kimi, DeepSeek, GLM, Qwen and others. It is updated regularly as new prompts are captured, and serves as reference material for AI red-teamers and researchers studying how vendors constrain model behavior and how those constraints can be extracted or bypassed.

github.comrepoPublished and added 3 Aug 2026🛡️ Security

A comprehensively curated landscape document tracking the offensive AI agent space: ranked open-source pentesting/red-team agent projects (e.g. strix, shannon, VulnBot), uncensored offensive LLMs (Qwythos, WhiteRabbitNeo/DeepHat, BaronLLM) versus security-specialized but aligned models (Foundation-Sec, CyberSecQwen), Claude Code Agent Skills and MCP servers wrapping tools like Metasploit and BloodHound, 116 cross-referenced academic papers on pentest/vuln-discovery agents, capability benchmarks (Cybench, NYU CTF Bench, CyberGym), and a fact-checked Top-20 list of commercial AI pentesting vendors. It also covers competitions like the DARPA AI Cyber Challenge and Tencent's Intelligent Penetration Hackathon, positioning itself as a cross-merge of several other awesome-lists in this niche.

github.comrepoPublished and added 1 Aug 2026🛡️ Security

Multi-Agent CAD (MAC) is a LangGraph-orchestrated multi-agent system that turns natural-language design briefs into editable engineering geometry (STEP/STL/GLB), decomposing mechanisms into separately generated, mated, and QA-validated parts and exporting URDF for robotics/physics simulation. It extends an earlier single-part Text-to-CAD workflow (~1,000 GitHub stars) with visual verification, an independent Judge agent, and an experimental assembly pipeline, reporting on a 10-prompt/141-feature benchmark that it matches a reproduced "CAD Skills" baseline's pass rate (140/141) while using tens-of-times fewer tokens and lower cost under Qwen-based configurations.

github.comrepoAdded 30 Sep 2026, published 30 Jul 2026🧠 AI

WanGP (Wan2GP) is an open-source 'super app' that packages dozens of open generative video, image, audio and TTS models (Wan, MiniMax H3, LTX-2, Qwen Image, Krea 2, SeedVR2, etc.) behind a single web UI, optimized to run on as little as 6GB VRAM including older Nvidia and AMD GPUs. It bundles LoRA support, quantized checkpoint formats (int8, fp8, GGUF, NVFP4), an agentic assistant ('Deepy') for orchestrating generation jobs, a plugin system, and a headless/API mode for batch and programmatic use.

github.comrepoAdded 22 Sep 2026, published 30 Jul 2026🧠 AI

whatisit is a CLI that turns plain-English requests into shell commands using a 941 MB Q4_K_M GGUF model (LoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct) served locally via llama.cpp on CPU. It can optionally use any OpenAI-compatible endpoint. It includes a denylist safety checker (DANGER/CAUTION) with 477 regression cases. On the 300-task InterCode-ALFA benchmark (NL2SH, NAACL 2025) the fine-tune scores 0.620 vs 0.540 for the base (McNemar p=0.004), roughly equal to the untuned 7B and about 11 points behind GPT-4o. A 3B variant is also available, and the author documents the training recipe, data mix and limits.

github.comrepoAdded 16 Oct 2026, published 29 Jul 2026🧠 AI

Better Harness is an MIT-licensed plugin/CLI from QoderAI that runs inside a coding agent (Claude Code, Codex, Qoder, Cursor, GitHub Copilot CLI, Qwen Code, Pi, Kimi Code and others) via a /better-harness command. It collects project and session evidence and scores five dimensions of the "Agent Work Loop" (feedforward guides like AGENTS.md, specs and Skills; feedback sensors like linters, tests and Hooks). It outputs HTML/Markdown/Canvas reports with prioritized findings, repair plans and acceptance checks, and keeps missing evidence explicit rather than inventing scores.

github.comrepoAdded 12 Oct 2026, published 21 Jul 2026🧠 AI

LLMVault is a self-hosted, deliberately-vulnerable training range modeled on WebGoat but for LLM applications, covering 25 labs across the OWASP Top 10 for LLM Applications (2025) - prompt injection, jailbreaking, data/model poisoning, agent exploitation, model extraction, and insecure downstream output handling. It ships two modes: a scored Play Mode with scripted assistants for repeatable flag capture, and an unscripted Live Mode that points the same attacks at a real local Ollama model (qwen2.5:3b-instruct) with a freshly minted per-session secret, requiring the player to actually talk the model into leaking it. Each lab pairs the attack with a written defense explanation, and an expert tier is shipped AES/Fernet-encrypted and unlocked manually by the maintainer.

github.comrepoPublished and added 18 Jul 2026🛡️ Security

Pentest Swarm AI is an open-source (AGPL-3.0), self-hosted CLI tool that runs dozens of LLM agents concurrently against a target, coordinating through a shared "stigmergic" blackboard rather than a fixed recon→exploit→report pipeline, so findings propagate and wake relevant agents in real time. It supports any provider with tool-calling (Claude, OpenAI-compatible endpoints, Together AI-hosted Llama/Qwen/DeepSeek/GLM, or fully local Ollama/LM Studio), ships bundled vulnerable labs (crAPI, Juice Shop, VAmPI, DVGA) for safe testing, and positions itself against the closed SaaS bug-bounty bot XBOW. It currently automates verified attack playbooks (BOLA/IDOR, mass assignment, NoSQL injection) with evidence-backed reporting, with roadmap items for runtime object-reference mining, self-correcting attacks, and on-demand specialist sub-agents.

github.comrepoPublished and added 21 Jun 2026🛡️ Security

A GitHub-hosted curated bibliography accompanying the survey "Advances and Challenges in Foundation Agents," organizing hundreds of papers across agent reasoning (CoT, ReAct, Tree/Graph of Thoughts), memory architectures (MemGPT, HippoRAG, Titans), multimodal foundation models (CLIP, LLaVA, Qwen2-VL), embodied/robotics agents (RT-1/RT-2, OpenVLA, Voyager), and alignment/RLHF techniques (DPO, SimPO, KTO, Constitutional AI). It functions as a living index into the foundation-agents research landscape, updated with new papers over time.

github.comrepoAdded 23 Sep 2026, published 7 May 2026🧠 AI