[AINews] High Taste Testers Ep. 1

What it claims

This AINews issue, covering 15-16 October 2026, centres on two US open-weight MoE releases. Meta's Muse Spark Open has 816B total and 42B active parameters. Reflection Beam 1.0 has 501B total and 23B active. Artificial Analysis (AA) scores them 39 and 30 on its Intelligence Index, first and second among US open-weight models. The Reddit recap adds releases the summary omits: NVIDIA Nemotron 3.5 Bolt 74B-A9B, a native 2-bit QAT model about 24GB on disk; Nemotron 3.5 Super VL 120B-A12B; StepFun Step-5-Preview at 600B/27B; and FLUX 3 Image Klein, a 5.6B image model. Most numbers come from tweets and threads, not papers.

How it works

Muse Spark Open has 62 layers and GQA with 16 KV heads. It activates 8 of 256 experts per token, uses a 1.8B Perception Encoder and has 1M context. It takes text, image, video and audio as input and outputs text.

Reflection Beam 1.0 uses sliding-window attention with a 512-token window in a 5:1 pattern. It has no RoPE, no shared experts, and uses gated attention, QK-norm, sandwich norm and logit softcapping (per @eliebakouch). It was trained on the open Mirror CLI harness.

Nemotron 3.5 Bolt is a Mamba-2/Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder and up to 256K context. It supports optional speculative decoding with 7 draft tokens.

Evidence

• Muse Spark Open uses 40k output tokens per AA task, the fewest among the top nine open models. It costs $0.84 per task, against $2.01 for GLM-5.3 and $0.27 for DeepSeek V4.1 Flash. It ranks last of those nine on AA's four agentic evals but leads open models on AA-Omniscience. Vals ranks it #12 among open-weight models and #38 overall. Vals also says input tokens drive most of its cost.
• Reflection Beam uses about 26k tokens per task. It scores 48.5% on AutomationBench-AA but 3.0% on Terminal-Bench 4.0. On AA-Omniscience it answers 33% correctly and hallucinates 47%. vLLM measured 1,769 output tok/s on 4× B300 with NVFP4 at TP4 and concurrency 16. SkyRL's +62% LoRA figure (276 to 447 tok/s/GPU) measures decode, not end-to-end training.
• Decision models: Liquid d1-ultra scores 65 on AA's Decision Index, ahead of Jev 1.14 at 57. Its non-hallucination rate is 85%, but accuracy falls to 32% from 41% for d1. It answers in 92ms and costs $0.042 per M input tokens. Perplexity pplx-decider v1.2 ships open weights at 27B (2¢/M) and 4B Flash (1¢/M).
• Hardware: SemiAnalysis reports TPUv7 at 1.42× the perf per dollar of GB300 NVL72 FP4, using a non-Google cost basis. GB200 FP4 keeps a 1.2× edge at low interactivity. Vera Rubin's 5.7× throughput-per-MW lead over GB300 narrows to 1.14-1.45× at peak throughput, and NVIDIA submitted all of those results. 8× MI355X beats HGX B300 by 24.4% in throughput on Kimi K3 at 50 tok/s per user.
• Haiku 5.5 costs $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.

How to use it

Muse Spark Open's BF16 weights are about 1.6TB. Meta's NVFP4 checkpoint quantizes only the experts and is about 530GB. Meta says it fits on 4 Blackwell Ultra GPUs, 8 H200s or two 512GB M5 Ultra Mac Studios. Together, Fireworks and Baseten charge $1.25/$4.25 per M input/output tokens, and $0.15 for cached input. Day-0 support exists in vLLM and mlx-lm (with vision). The Reddit post also lists SGLang, TensorRT-LLM, llama.cpp and Unsloth.

Reflection Beam ships under Apache 2.0 with BF16, FP8 and NVFP4 weights, a DFlash2 drafter and a safety report. 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. Nemotron 3.5 Super VL ships under OpenMDW-1.1. NVIDIA recommends its NVFP4 variant, with examples on 4× H100 or 2× GB200. Step-5-Preview recommends temperature 1.0, top_p 0.95 and High reasoning mode. Its Q3_K_M quant is reportedly 286GB.

Gemini 4 Argon now allows 1M output tokens, up from 64K, shared with the input window. Intro pricing is $2/$10 until 30 November 2026.

Limits

• Muse licence: The Muse Model License 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.
• Muse safety tuning: Reddit commenters say safety post-training blocks cybersecurity work. One claims the model scores below Muse Spark 1.2 on benchmarks.
• Step-5-Preview licence: It is non-commercial, unlike earlier StepFun flash models under Apache 2.0. One commenter claims it is byte-identical to a leak from about 25 days earlier.
• China export rules: This is a draft from MOFCOM and MOST, with public comments open until 15 November 2026. Temporary pre-export registration applies immediately. Already-open-sourced models are exempt. Thresholds are reported as 2T parameters or 10^26 FLOPs, but the Reddit summary calls them undefined.
• Interpretability: The Qwen3-32B neuron audit is preliminary. Stella Biderman (@BlancheMinerva) calls "interp is solved" claims hype.
• Unconfirmed items: Codex model rumours and the Anthropic $1.5T valuation are unconfirmed reports.