Skip to content
Main Site News Console

[AINews] The Discerning Tester: Issue 1

· Latent Space Translated
播客深度访谈

We are piloting a new show on Latent Space: High Taste Testers, a spinoff of our food show that brings you right into the vibe of a typical San Francisco lunch conversation. From DevDay to the latest models and assistants, covering everything you care about most:

Let us know what you think!

AI News for October 15–16, 2026. We reviewed 12 subreddits, 544 Twitter accounts, and no extra Discords. You can search all past issues on the AINews website. Friendly reminder: AINews is now a section on Latent Space.


AI Twitter Highlights

US Open-Weight Models Surge: Muse Spark Open and Reflection Beam

  • Muse Spark Open (Meta): An 816B total, 42B active parameter open-weight Mixture-of-Experts (MoE) model. Artificial Analysis scored it a 39 on their Intelligence Index—the highest score among US open-weight models, beating NVIDIA’s Nemotron 3.5 Super VL (26) by 13 points (AA).

    • Architecture & Specs: 62 layers, GQA (Grouped-Query Attention) with 16 KV heads, 8 of 256 experts activated per token, a 1.8B Perception Encoder, and 1M context support. Inputs support text, image, video, and audio; output is text. BF16 weights are around 1.6TB. Meta’s NVFP4 checkpoint quantizes only the experts, reducing the weight size to ~530GB. Meta states the model can run on 4x Blackwell Ultra GPUs, 8x H200s, or two 512GB M5 Ultra Mac Studios.
    • Pricing & Efficiency: Together, Fireworks, and Baseten charge $1.25 / $4.25 per million input/output tokens, with cached input at $0.15. AA estimates its cost at $0.84 per benchmark task, compared to $2.01 for GLM-5.3 and $0.27 for DeepSeek V4.1 Flash. It uses 40k output tokens per task—the fewest among the top nine open models (Cost). Vals reported it used 55.5M total output tokens across their benchmark suite—the lowest among all open models—though noted input tokens account for the vast majority of its cost (Vals).
    • Weaknesses: Across AA’s four agentic evals, it ranked dead last among the top nine open models, though it led open models on AA-Omniscience. Vals ranks it 12th among open-weight models and 38th overall (Vals). Arena ranked it as the #1 US open model on Code Arena WebDev (score 1511), and #13 among all open models (Arena).
    • License: The proprietary Muse Model License requires attribution: “Built with Muse Spark Open”. Model-as-a-Service (MaaS) providers with over $20M in annual revenue must negotiate a separate agreement with Meta.
    • Ecosystem Support: mlx-lm had day-one support with vision compatibility, and an official Apple blog post detailed distributed inference across Mac Studios via RDMA (@zcbenz). vLLM powered the RL rollouts during Muse’s training phase and supported the model out of the box at launch (vLLM). SemiAnalysis has added B200 configs in InferenceX (SemiAnalysis).
    • Criticism (Hot Takes): @teortaxesTex called the model “dead on arrival (DOA)” compared to the smaller, MIT-licensed V4.1 Flash. He also baselessly speculated that Meta held back Spark 1.3 due to distillation from OpenAI (Speculation). Meanwhile, @alexandr_wang remarked that MSL has built a “machine for making new models.”
  • Reflection Beam 1.0: A 501B total, 23B active parameter MoE model with 1M context, released under the Apache 2.0 license. The release includes BF16, FP8, and NVFP4 weights, a DFlash2 drafter, the open-source Mirror CLI harness used to train Beam, and a safety report (Reflection, Safety Report, Mirror).

    • Independent Evals: AA scored it 30, ranking #2 among US open models, using ~26k tokens per task. It achieved 48.5% accuracy on AutomationBench-AA, but only 3.0% on Terminal-Bench 4.0. On AA-Omniscience, it had a 33% correct rate alongside a 47% hallucination rate (AA).