AI Daily · Oct 16, 2026 · Friday
BAD INFLUENCE
Meta releases Muse Spark Open, a 816B open-weight model for coding and agents
Meta says it is releasing Muse Spark Open, which it calls the most capable US open-weight model to date. The model is built for coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning. It has 816B total parameters with 42B active, a 1M token context window, and native support for text, image, video, and audio.
- Meta releases Muse Spark Open with API access and downloadable weightsAI at Meta
- Alexandr Wang announces Muse Spark Open API and weight availabilityAlexandr Wang
- vLLM adds day 0 support for Meta's Muse Spark Open, a 816B open-weight MoE modelvLLM Blog
- SGLang shows how to run Muse Spark Open for agentic tasksSGLang
Model releases & updates
NVIDIA releases Nemotron 3.5 Super VL, an open 120B multimodal agent model
NVIDIA has released Nemotron 3.5 Super VL, a 120B-parameter mixture-of-experts model that activates 12B parameters per token and handles text, images, documents, and video. The weights are available under the OpenMDW-1.1 license with a context window of up to 1M tokens. NVIDIA reports 89.90 on GPQA-Diamond and 17.78 on tau3-bench banking, using self-reported scores from its NVFP4 checkpoint.
Comfy BlogOfficial5 more sourcesFLUX 3 Image Klein arrives in ComfyUI as a 5.6B open-weights model
Black Forest Labs' FLUX 3 Image Klein is a 5.6B-parameter open-weights image model now available in ComfyUI, in base and distilled (turbo) forms. Paired with INT8 ConvRot quantization and the distilled model, 1024×1024 text-to-image generation runs 4.3× to 9.3× faster than the BF16 50-step baseline, with the largest gains on consumer GPUs. On an RTX 5070, one image drops from 71 seconds to under 8.
- Black Forest Labs releases FLUX 3 Image Klein, optimized for NVIDIA GPUsBlack Forest Labs
- Hugging Face introduces FLUX.3 Image Klein in DiffusersHugging Face Blog
ARC Prize@arcprizeOfficial3 more sourcesARC Prize verifies Gemini 4 Argon scores above 95% on ARC-AGI-3
ARC Prize reports verified ARC-AGI results for Google's Gemini 4 Argon. It scores 97.5% on ARC-AGI-3 with a provider adapter harness at $4.8K, and 49.8% at $5.8K with the standard harness. It also reports 91.3% on ARC-AGI-2 at $0.32 per task and 98.5% on ARC-AGI-1 at $0.16 per task. The post says Google DeepMind is the second frontier AI lab to score above 95% on ARC-AGI-3.
Mistral AI releases Timestral, a family of time-series foundation models
Mistral AI releases Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. The largest model ranks first on GIFT-Eval, TIME and fev-bench among pre-trained zero-shot models, and the 209M model runs in about 140 ms per forecast on a standard server CPU. Timestral is available on-prem, in a VPC or air-gapped, with fine-tuning through Mistral Forge, under a modified MIT license.
Product launches & updates
Meta explains how it built the Forum app with AI-native engineering
Meta's engineers built Forum, a standalone Groups app, by bringing an AI agent into the development cycle. They say the agent handled planning, building, testing and verification, with engineers setting goals and approving results. Meta reports that Forum reached core feature parity with Facebook Groups in eight weeks, and that an agent-run stress test found a memory issue that was then fixed before launch, roughly halving memory use and page faults.
NVIDIA Vera Rubin NVL72 beats GB300 NVL72 on AA-AgentPerf throughput per MW
Artificial Analysis reports its first AA-AgentPerf results for NVIDIA's Vera Rubin NVL72 rack. At peak throughput, it serves 1.14x to 1.45x more tokens per MW than GB300 NVL72 across four models, with leads up to about 5.7x at higher per-user speeds. The rack's 2.4x memory bandwidth and 2.3x dense NVFP4 compute per GPU are NVIDIA's stated specifications.
Industry news
Anthropic NewsroomOfficial5 more sourcesAnthropic gives METR employee-like access to evaluate its frontier models
Anthropic says it has onboarded six METR employees with company hardware, office space, and internal access roughly comparable to its own risk assessment teams. METR will publish its analysis of Anthropic's risk levels, incidents, or practices, and Anthropic can require redaction only of narrow categories such as customer data, privileged information, and model architectures. Anthropic also describes three access tiers for other evaluators, including SecureBio, which is currently in the first tier reviewing its chemical and biological weapons chapter.
Papers & research
Shanghai AI Lab's Shusheng-S2 helps prove a three-dimensional Ghys conjecture result
Shanghai AI Laboratory and Professor Xu Disheng's team say the Shusheng-S2 scientific multimodal model, working with human researchers, produced a proof of the three-dimensional Ghys conjecture. Peking University's Dong Bin and Liu Ruochuan team says it completed a Lean formal verification of all four main theorems in about five days, producing roughly 550,000 lines of Lean code. The collaboration ran seven rounds of recursive reasoning, growing the problem graph from 15 to 389 nodes.
Tips & takes
Google Developers BlogOfficialGoogle's adk-recipes repo organizes agent examples into three layers
Google Developers Blog describes google/adk-recipes, which splits content into core/ reference patterns, plugins/ installable vertical skills, and contrib/ deployable services. The post walks through the Horizon reference harness and the retail Plugin Recipe, with setup and install commands for each layer.
JetBrains AI BlogOfficialNetflix benchmarks JetBrains IDE intelligence in coding agents, reporting 20–30% efficiency gains
Netflix integrated a headless JetBrains IDE code-intelligence engine into Claude Code and Codex and benchmarked it on two Java and Kotlin codebases. Across models and tasks, agents typically used 20–30% fewer tokens and less time, with gains up to about 65% in the best case. Netflix built the daemon and CLI needed to make the LSP callable by agents, because JetBrains had not yet shipped one.
Apple shows how to run an 816B-parameter model across Mac Studios with MLX and RDMA
Apple's machine learning research team says MLX can run Muse Spark Open, an 816B-parameter mixture-of-experts model, across four Mac Studios with M5 Ultra at over 36 tokens per second. The post compares pipeline and tensor parallelization over Thunderbolt RDMA and gives the mlx.launch commands. It also reports TTFT and decode speedups over the M3 Ultra for Qwen 3.8 and Mistral Medium 3.5.