[AINews] High Taste Testers Ep. 1

What it claims

This issue covers a dense one-day news cycle (10/15–10/16/2026). For an engineer, the useful parts are the deployment specifics behind the headline releases: footprints, licence terms, serving numbers and independent eval placements. Several of these change the adopt/skip decision:

• Muse Spark Open fits on 4 Blackwell Ultra GPUs in NVFP4, but it ranks last among the top nine open models on agentic evals.
• Reflection Beam 1.0 is Apache 2.0, but its Terminal-Bench 4.0 score is near zero.
• Step 5 Preview is non-commercial.
• Nemotron 3.5 Bolt squeezes a 74B model into 24GB via native 2-bit QAT.

How it works

• Muse Spark Open (816B/42B-active):
• 62 layers, GQA with 16 KV heads, 8 of 256 experts active per token.
• 1.8B Perception Encoder, 1M context.
• Text, image, video and audio in; text out.
• Reflection Beam 1.0 (501B/23B-active):
• Sliding-window attention with a 512 window in a 5:1 pattern, and no RoPE.
• Llama-4-style global attention scaling, gated attention, QK-norm, sandwich norm, logit softcapping, and no shared experts.
• Nemotron 3.5 Bolt 74B-A9B: a Mamba-2–Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder, native IQ2 2-bit QAT and up to 256K context.
• Nemotron 3.5 Super-VL 120B-A12B: interleaved Mamba-2, attention and MoE layers, 1M context and native MTP speculative decoding.
• FLUX 3 Image Klein:
• A distilled 5.6B image transformer with Qwen3-VL-4B-Instruct as its text encoder.
• Expects full-scene captions, often produced by the flux-3-image-klein-pup upsampler.
• Supports bounding-box layout control on a 0–1000 grid.

Evidence

• Muse Spark Open, quality:
• Artificial Analysis (AA) Intelligence Index 39, against 26 for Nemotron 3.5 Super VL.
• Last of the top nine open models on AA's four agentic evals, while leading open models on AA-Omniscience.
• Vals ranks it #12 among open-weight models and #38 overall.
• Code Arena WebDev 1511, the top US open model.
• Muse Spark Open, cost:
• $0.84 per AA index task, against $2.01 for GLM-5.3 and $0.27 for DeepSeek V4.1 Flash.
• 40k output tokens per task, the fewest among the top nine open models. Vals says input tokens drive most of its cost.
• Reflection Beam:
• AA index 30, at about 26k tokens per task.
• AutomationBench-AA 48.5%, but Terminal-Bench 4.0 only 3.0%.
• On AA-Omniscience it answers 33% correctly and hallucinates at 47%.
• vLLM measured 1,769 output tok/s on 4× B300, NVFP4, TP4, at concurrency 16.
• SkyRL reports +62% LoRA decode throughput (276→447 tok/s/GPU). This is a decode figure, not an end-to-end training speedup.
• Haiku 5.5: $0.10/$0.50 for prompts up to 100K tokens. It scores 39.2% on Terminal-Bench 4.0 at max effort, against 16.4% for GPT-6 Luna.
• Liquid d1-ultra:
• AA Decision Index 65, against 57 for Jev 1.14.
• Non-hallucination rate rises to 85% from 2% for d1, but accuracy falls to 32% from 41%.
• 92ms per task, $0.042 per M input tokens, output free.
• Hardware:
• TPUv7 has 1.42× better perf per dollar than GB300 NVL72 FP4, on an external cost basis, and 2.5× versus MI355X. GB200 FP4 keeps a 1.2× edge at low interactivity.
• Vera Rubin's lead over GB300 is up to 5.7× throughput per MW, narrowing to 1.14–1.45× at peak throughput. NVIDIA submitted all of these results.
• 8× MI355X gives 24.4% more throughput than HGX B300 on Kimi K3 at 50 tok/s per user.
• Agent research:
• Forced forgetting (resetting agents to the best candidate so far) gives up to 15% over parallel agents and 39% over a single agent at equal budget.
• On FDE-Gym, the best agent fully resolves only 14 of 95 real customer-delivery tasks.
• OpenHands' apparent stability comes mainly from hardcoding temperature 0.

How to use it

• Muse Spark Open, weights:
• BF16 is about 1.6TB.
• Meta's NVFP4 checkpoint quantizes only the experts and comes to about 530GB. Meta says it fits on 4 Blackwell Ultra GPUs, 8 H200s or two 512GB M5 Ultra Mac Studios.
• Muse Spark Open, serving:
• Together, Fireworks and Baseten charge $1.25/$4.25 per M input/output tokens, with $0.15 cached input.
• Day-0 support in vLLM and mlx-lm (with vision). Also listed: SGLang, TensorRT-LLM, llama.cpp and Unsloth.
• Reflection Beam:
• Ships BF16, FP8 and NVFP4 weights, a DFlash2 drafter, and the Mirror CLI harness it was trained on.
• SkyRL supports LoRA, full fine-tuning and MXFP8 RL.
• Ollama serves it free for a limited time, US-hosted with zero data retention.
• Nemotron 3.5 Bolt: targets 24GB local deployment with vLLM on Ampere, Hopper or Blackwell, with optional speculative decoding using 7 draft tokens.
• Nemotron 3.5 Super-VL:
• Deployment examples use 4× H100 or 2× GB200.
• NVIDIA recommends the NVFP4 variant for inference.
• Licence: OpenMDW-1.1.
• Step 5 Preview (600B/27B-active):
• Recommended settings: temperature=1.0, top_p=0.95, High reasoning mode.
• The Q3_K_M build is reportedly about 286GB.
• pplx-decider v1.2: open weights at 27B (2¢/M) and a 4B Flash model (1¢/M). Adds reranking and single-pass multi-question inference that is up to 5× faster.
• FLUX 3 Klein:
• Ships BF16, FP8 and NVFP4 checkpoints under the FLUX Kommunity License.
• ComfyUI weights are at black-forest-labs/flux-3-image-klein-comfy; the ComfyUI loader is only in the latest git version.
• cc-stretch (MIT) is a Claude Code plugin that compresses replies and lets you expand them interactively:

/plugin install cc-stretch --marketplace pro-vi/cc-stretch
/tui fullscreen

Limits

• Licences:
• Muse requires "Built with Muse Spark Open" attribution.
• Model-as-a-Service businesses with more than $20M in annual revenue need a separate agreement with Meta.
• Step 5 Preview is non-commercial, unlike StepFun's earlier Apache 2.0 flash models. A commenter claims it is byte-identical to a leak from about 25 days earlier.
• Safety tuning: Reddit users report that Muse is heavily safety-tuned, which they say blocks security research, and some claim it benchmarks below Muse Spark 1.2.
• FLUX 3 Klein: early testers report anatomy errors, weak world knowledge and heavy censorship.
• Vendor-sourced numbers: the Vera Rubin figures are NVIDIA-submitted. The TPUv7 cost basis is an estimate, not Google pricing.
• Unverified claims: the Qwen3-32B neuron audit (about 90% of neurons with one main job) is explicitly preliminary and disputed. Tencent's Hy-Memory claim against Opus 5 and Elorian's Glimpse results are self-reported.
• Pending policy: China's export-control draft is not yet in force. Comments run until 2026-11-15, and its thresholds are undefined.