Open weights go frontier-class — a 501B base model, a 816B Meta MoE, and Gemini 4 Argon land in one week

This digest covers model releases and updates from 2026-10-10 through 2026-10-17, scanning Anthropic, OpenAI, Google DeepMind, Meta AI, Mistral, Hugging Face, and Simon Willison's blog.

🔥 Highlights

1. Gemini 4 Argon: Built to solve your hardest problems — Google's new frontier model goes GA with a 1M-token unified context.
   

2. Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta — an 816B-parameter open-weight model Meta calls the most capable US open model yet.
   

3. Welcome Beam: a 501B open model you can own, customize, and scale — a brand-new lab's first open-weight frontier-scale release.
   

4. Release: llm-meta-ai 0.2 — covers Meta's closed Muse Wonder, possibly its first "Fable class" model.
   

5. Introducing Timestral — Mistral's first foray into time-series forecasting, running CPU-only.
   

Anthropic

Nothing relevant this week. The news listing had only policy and business-expansion posts (an update on independent evaluators, midterm-election prep, a new Singapore office) — no model launches, version updates, or capability/pricing changes.

OpenAI

• Introducing the Rosalind Academic Plan (2026-10-15): OpenAI is giving eligible U.S. academic research labs access to GPT-Rosalind, its dedicated life-sciences model (frontier reasoning plus the Rosalind Workbench for medicinal chemistry, genomics, and wet-lab tool orchestration), bundled with up to 30 Business Standard seats free for up to 12 months. It's a new-availability and pricing change rather than a new model — note that API access is explicitly excluded, so it's chat/workbench-only for now.

Google DeepMind

• Gemini 4 Argon: Built to solve your hardest problems (2026-10-13): Google shipped Gemini 4 Argon to general availability, targeting complex multi-step work — coding, finance, legal, cybersecurity, science — with a 1M-token unified input/output context. Introductory API pricing is $2/M input and $10/M output tokens (cached input 95% off) through November 30, 2026, after which it doubles to $4/M input and $20/M output; it's live in the Gemini app, AI Studio, Android Studio, Antigravity, and Gemini for Business.
• Gemini 4 Argon is now available for developers (2026-10-13): Companion developer post confirming the model is live on the paid Gemini API tier for long-horizon engineering tasks like large code migrations, with AI Ultra subscribers also getting early access via Antigravity.

Meta AI

Nothing relevant this week on ai.meta.com/blog — its newest post is still from July 2026. (Meta's actual model news this week, Muse Wonder and Muse Spark Open, published on separate properties and is covered above and in the Hugging Face / Simon Willison sections.)

Mistral

• Introducing Timestral (2026-10-15): Mistral's first time-series foundation model family, in four sizes (5M, 28M, 123M, 209M parameters) for probabilistic, multivariate forecasting — sensor monitoring, anomaly detection, capacity planning. Mistral claims the 209M model is #1 among zero-shot pretrained models on GIFT-Eval/TIME/fev-bench; it runs CPU-only (10–140ms per forecast), supports a 32,768-step context, and ships under a "Modified MIT" license with weights on GitHub and Hugging Face — practical for teams that want forecasting without standing up a GPU.

Hugging Face

• Welcome Beam: a 501B open model you can own, customize, and scale (2026-10-16): Reflection, a new US lab, launched Beam — a 501B-parameter open MoE (23B active params/token, Apache 2.0) with a 1M-token context window aimed at coding and agentic work. Day-0 support lands in transformers, vLLM, and SGLang with SFT/RL recipes; an NVFP4 checkpoint (~280GB) runs on 4 B200 GPUs, while full BF16 needs ~1TB.
• Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta (2026-10-15): Meta's open-weight multimodal MoE (~816B total / ~42B active params, up to 1M-token context) is pitched as the most capable US open-weights model to date, tuned for agentic coding. It takes text, image, audio, and video input (text-only output); quantized builds run on high-memory Macs and via llama.cpp, vLLM, SGLang, MLX, and HF Inference Endpoints (BF16 is ~1.5TB).
• Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint (2026-10-15): IBM's modular Granite Switch line (8B, 30B) now bundles a base chat model with a Decision adapter, a Guardian safety adapter, and an embedded speech-to-text model behind a single endpoint. Because the adapters reuse the base model's KV cache, the full bundle costs about the same GPU memory as the base model alone — useful if you need chat, structured decisions, and transcription without running three separate deployments.
• Welcome FLUX.3 Image Klein in Diffusers (2026-10-15): Black Forest Labs' FLUX.3 Image Klein gets day-one Diffusers support, adding text-to-image, multi-reference image-to-image, and bounding-box layout control for placing image regions. The distilled checkpoint generates a 1024x1024 image in ~2 seconds on an H100; quantized FP8/NVFP4 weights and LoRA training scripts let it run on smaller GPUs.
• Introducing Lumma-Fev-27B (2026-10-14): FrontiersMind's 27B addition to its Lumma-Fev "decision model" family scores supplied options in a single forward pass across English, Hindi, Tamil, and Kannada. It beats TypeSafe Jev on held-out decision accuracy (90.7% vs 83.9%) with better calibration, though Jev still leads on general reasoning; weights are Apache 2.0, needing ~49GiB in bfloat16 (one 80GB GPU).
• HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost (2026-10-14): Ai2's updated climate emulator adds a downscaling model that produces several high-resolution fields at once and more accurately reproduces extreme events like tropical cyclones. Paired with an updated coarse ACE model, it generates global 3km climate data at roughly 300x lower cost than the X-SHiELD simulator it was trained on; weights are on Hugging Face.
• UltraBERT: One Encoder for Classification and Retrieval (2026-10-14): DMIS Lab's new English encoder family (150M/380M params, 32K context) plus an UltraBERT-Embed-150M variant is trained to handle both classification and retrieval from one encoder. The 380M model scores 91.0 on GLUE and 47.7 on BEIR; the 150M embedder is the top dense model under 200M params on MTEB(eng, v2) — all weights Apache 2.0.
• Run fast, private agents on-device with LFM2.5-24B-A2B (2026-10-13): Liquid AI's 24B reasoning model activates only 2B params/token for on-device agents. Versus its predecessor it extends context from 32K to 131K tokens, doubles vocabulary to 128K, and adds a multi-stage agentic post-training pipeline; it matches or beats the larger Qwen3.6-35B-A3B on several tool-use benchmarks and runs at 184 tokens/s on an Apple M5 Max.
• ColGrep-Agent: Cutting Frontier Coding Agents' Tokens by a Third (2026-10-12): LightOn's open 2B-parameter local code-search model lets coding agents delegate repository search instead of using grep. Tested with Claude Code, delegating search cut API tokens by 34% and cache-aware cost by 40% with no significant accuracy loss, at the cost of higher latency (13.0s to 21.6s median); GGUF weights are available now.
• UIBK-Embedding-v1: A Dense and a Late-Interaction Retriever in One Model, One Index and One Search (2026-10-12): A 179M-parameter English retrieval model (Apache 2.0) produces both a dense embedding and token-level late-interaction vectors from a single encoder, reaching 59.79 nDCG@10 on BEIR. It lets teams replace separate dense + ColBERT hybrid setups with one model and one index, loadable as a standard ColBERT model in PyLate.

Simon Willison

• Release: llm-gemini 0.35 (2026-10-13): Willison's llm-gemini plugin adds support for Gemini 4 Argon via the Gemini API and the $100/month subscriber plan. Launch pricing is $2/M input and $10/M output tokens, but it's set to double on December 1st, putting it at parity with Claude Opus 5.5's pricing — worth locking in usage now if cost matters.
• Release: llm-meta-ai 0.2 (2026-10-12): Willison's llm-meta-ai plugin adds support for Meta's newly released "most capable models," Muse Wonder and Muse Wonder Full, which he speculates may be Meta's first "Fable class" models. Artificial Analysis rates Muse Wonder (Max) roughly on par with GPT-6.1 Sol; the models expose six reasoning levels (minimal through max), continuing Meta's pattern of roughly monthly releases.

Through-line

The frontier-vs-open-weights line kept blurring this week: Google pushed a closed, pay-per-token flagship (Gemini 4 Argon) to GA at the same time two separate open releases — Meta's own 816B Muse Spark Open and newcomer Reflection's 501B Beam — landed within a day of each other, both explicitly marketed as agentic-coding models with million-token context. Meanwhile smaller, task-specific releases (Mistral's CPU-only Timestral, LightOn's grep-replacing ColGrep-Agent, IBM's adapter-bundled Granite Switch) suggest teams are increasingly shipping narrow, cheap-to-run models alongside the giant general ones rather than waiting for a single model to do everything.

What caught your eye this week — the open-weights race, or the pricing moves? Drop a comment below.