Minerka The AI briefing

Library

Models

The systems themselves: assistants, open weights, and specialists. Name the layer before you repeat the claim.

How models are ranked, and what they cost · Start here

Follow this shelf

Worked check: Carbon’s introductory list is $2/$10 per million tokens — not on the public API yet; DeepMind’s Index v5 claim of 59 is DeepMind’s report. Test named. Price named. Door separate.

Start here

One worked check, four list prices, then three questions — before you trust what changed this week.

Worked check: PlanAssist cut officer assessment time up to 58% in pilots — DeepMind’s figure; a human still decides. Field card · prices and 4 checks

Pocket list (November 1, 2026)

  1. Measurement — which named test, on which date, scored by whom.
  2. Bill — list input and output per million tokens on the vendor page.
  3. Door — live API, lab claim, replay, or a downloadable file — keep them separate.

Models · Agents · Fine-tuning · All chapters

In this briefing

  1. Mistral releases the weights for Large 4 (Le Chonk), a 1T-parameter open modelHigh impact · October 29, 2026 · On October 29, 2026, Mistral released open weights for Mistral Large 4, its 1-trillion-parameter multimodal Mixture-of-Experts model with about 49 billion active parameters per token. Mistral says the model scores 41 percent on Terminal-Bench 4.0 and 48 percent on the Artificial Analysis Coding Agent Index while using 47 percent fewer tokens than GLM-5.3 on that index.
  2. A claimed ~14T Anthropic “numbat” checkpoint on Hugging Face has 43 tensors that match a name-seeded RNGMedium impact · November 1, 2026 · A Hugging Face upload, burner22/numbat, claims about 14 trillion parameters as a mid-pretraining Anthropic Mythos-class checkpoint. Hugging Face’s automatic count shows about 6.8T. On November 1, 2026, a community audit on the repo’s discussion #13 reported that 43 complete tensors — 730,368 BF16 values — regenerate bit-for-bit from a NumPy PCG64 recipe seeded by each tensor’s name, at revision 6b7f8fc. Anthropic is not a publisher on the card.
  3. Gemini 4 Carbon lists at $2/$10 — not on the public API yet; DeepMind claims Index v5 lead at 59High impact · October 28, 2026 · On October 28, 2026, Google DeepMind announced Gemini 4 Carbon. DeepMind says it ranks first on Artificial Analysis Intelligence Index v5 at 59. The introductory price is $2 per million input tokens and $10 per million output tokens. It is not on the public API yet.
  4. NVIDIA posts an open transcription model, 4 billion parameters active, at 5.33 percent word errorMedium impact · October 28, 2026 · On October 28, 2026, NVIDIA released Nemotron 3.5 Transcribe, an open-weight speech model. About 4 billion of its 30 billion parameters are active. On the Hugging Face English leaderboard, among open-weight models, the average word error is 5.33 percent. The post does not name the license.
  5. Ai2 finds a novelty judge can fall from 93.5 percent to 40.9 percent when the prompt changesMedium impact · October 27, 2026 · On October 27, 2026, the Allen Institute for AI released a study of models that score whether a research idea is new. For GPT-5.4, one prompt scored 93.5 percent on pairs of human and generated ideas. A shorter prompt on the same pairs scored 40.9 percent. A simple Claude Opus 4.6 prompt cost $2.59 and beat two purpose-built judges that cost $30.66 and $82.05.
  6. CUDA 13.4 can refuse a GPU allocation that crosses a memory limitMedium impact · October 27, 2026 · On October 27, 2026, NVIDIA said CUDA 13.4 enforces a per-workload GPU memory boundary. A job can ask for an exact size, and an allocation past that size fails. KAI Scheduler v0.18.0 can write that limit from Kubernetes. Time-slicing keeps each job in its own error domain. The parallel mode does not.
  7. Liquid AI ships LFM3, a 1.7 billion parameter model that beats its larger predecessorHigh impact · October 27, 2026 · On October 27, 2026, Liquid AI released LFM3, six models in the 2 billion parameter range on one backbone. Liquid says LFM3-1.7B beats LFM2.5-2.6B with 35 percent fewer parameters. On WebArena-Verified the score is 38.42, against 3.02 for the predecessor. The weights are on Hugging Face. The post does not name the license.
  8. DeepMind shows Genie 4 to trusted testers, with Gemini 3.8 Flash steering the worldMedium impact · October 27, 2026 · On October 27, 2026, Google DeepMind announced Genie 4, a world model that generates a scene and its sound together. It is a research preview for trusted testers. DeepMind says Gemini 3.8 Flash watches the frames and checks that each step happened. The post does not give a frame rate or a resolution, and the model is not in the public API.
  9. NVIDIA says the 4 GB Nemotron Bolt model is available, and uses fewer tokensMedium impact · October 27, 2026 · On October 27, 2026, NVIDIA said Nemotron 3.5 Bolt E4B, the 4 GB model, is available for on-device game characters. NVIDIA says it is about 40 percent higher quality than its prior model on the benchmarks it chose, and that it writes up to about 80 percent fewer tokens than unnamed peers. The context window is 64,000 tokens.
  10. GPT-6.1 Terra is $1 and $5 per million tokens, and OpenAI's coding score more than doublesHigh impact · October 26, 2026 · OpenAI released GPT-6.1 Terra on October 26. Standard API prices are $1 per million input tokens, $0.05 cached, and $5 per million output tokens, which OpenAI says is half the price of GPT-5.6 Terra and GPT-6.1 Sol. On OpenAI's Terminal-Bench 4.0 row, the max-effort score is 50.3 percent, against 21.5 percent for GPT-5.6 Terra.
  11. Anthropic filed a public S-1 and named the ticker ANTMedium impact · October 26, 2026 · On October 26, 2026, Anthropic filed a Form S-1 for a proposed initial public offering of Class A common stock. It intends to list on the Nasdaq Global Select Market under the symbol ANT. The number of shares and the price range are not set, and the registration is not yet effective.
  12. Kimi K3.1 Pro posts 65.7 percent on Terminal-Bench, with weights due by November 4High impact · October 24, 2026 · Moonshot released Kimi K3.1 Pro on October 24, a post-trained Kimi K3 it calls a 3-trillion-class model, with a 1 million token context. On Moonshot's Terminal-Bench 4.0 row it scores 65.7 percent, ahead of GPT-6 Astra at 58.2 and just behind Claude Opus 5.5 at 66.4. The weights are not on the page. Moonshot says they will be released by November 4.
  13. Epoch grades two leading models at 65 percent of its own work standardMedium impact · October 8, 2026 · Epoch AI scored models on tasks from its own work. Claude Fable 5.1 and GPT-6 Astra each averaged 65 percent of the rubric. Epoch says 100 percent is the line at which a task can be handed over, and that a 65 percent result is still a failed task.
  14. Mirendil says one researcher pretrained a coding model for under $10 millionHigh impact · October 21, 2026 · On October 21, 2026, Mirendil said one researcher, working with agents, pretrained Mirendil-Base in 3.5 months for less than $10 million in compute. The post puts the model in the 100 to 500 billion parameter range.
  15. Mistral prices Voxtral 3, including a 4B transcription model at $0.0035 a minuteHigh impact · October 20, 2026 · On October 20, 2026, Mistral released Voxtral 3, speech models plus a voice-agent SDK. Voxtral Mini Transcribe 3 is 4 billion parameters at $0.0035 a minute.
  16. MiniMax posts M3.1-Flash, about 428B parameters with 23B activeHigh impact · October 17, 2026 · The MiniMax-M3.1-Flash card lists about 428 billion parameters, 23 billion active per token, and a 1 million token context. The license name on the card is MiniMax Community.
  17. DeepSeek posts a 763B MIT checkpoint, V4.1-Flash-1017High impact · October 17, 2026 · On October 17, 2026, DeepSeek shipped V4.1-Flash-1017 on its website, app, and API under the name deepseek-flash, and published MIT weights. The card lists 763 billion parameters and a 1 million token context.
  18. Mistral releases Timestral, a 209M model for forecastsHigh impact · October 15, 2026 · On October 15, 2026, Mistral released Timestral in four sizes, from 5 million to 209 million parameters. The license on the post is a modified MIT.
  19. Reflection releases Beam, a 501B open model under Apache 2.0High impact · October 16, 2026 · On October 16, Reflection published Beam's weights. The model has 501 billion parameters, about 23 billion active, and a 1 million token context. The license is Apache 2.0.
  20. Meta releases Muse Spark Open, an 816B model that runs on two Mac StudiosHigh impact · October 15, 2026 · Meta published the weights for Muse Spark Open on October 15. It has 816 billion parameters, 42 billion active, and a 1 million token context. Meta says two M5 Ultra Mac Studios can hold it.
  21. Gemini 4 Argon opens to developers at $2 per million input tokens until November 30High impact · October 13, 2026 · Google made Gemini 4 Argon generally available on October 13. The introductory API price is $2 per million input tokens and $10 per million output tokens. On December 1 that becomes $4 and $20.
  22. Microsoft-Decision-1 scores fixed choices, and Microsoft says it is 35 times quicker than GPT-6 SolHigh impact · October 9, 2026 · On October 9, Microsoft released a model that returns a score for each allowed answer instead of writing text. On Microsoft's tests it was the fastest of the set, 35 times quicker than GPT-6 Sol.
  23. Meta releases Muse Wonder for agents that remember and keep goingHigh impact · October 12, 2026 · Meta launched Muse Wonder and Muse Wonder Full on October 12. Developers can call Muse Wonder now. The fuller model is for Muse Code subscribers.
  24. GPT-6 arrives in ChatGPT, and it can draw the screenHigh impact · October 7, 2026 · OpenAI is rolling GPT-6 into ChatGPT, and a reply can include buttons, forms, charts, and small tools.
  25. EmbeddingGemma 2 searches text, images, audio, and video on the deviceHigh impact · October 6, 2026 · Google DeepMind released a 740-million-parameter open model that embeds text, images, audio, and video on a device.
  26. NVIDIA releases Nemotron 3.5 Super VL, a 120B open modelMedium impact · October 15, 2026 · On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open model with 120 billion parameters, and a 24 GB edge model it calls Bolt 74B-A9B.
  27. Liquid AI releases a 24B agent model with 2B activeMedium impact · October 13, 2026 · On October 13, 2026, Liquid AI released LFM2.5-24B-A2B, a 24 billion parameter model with 2 billion active and a 131,000 token context. The card's license name is LFM 1.0.
  28. Google reports 87 percent for Argon on a robotics reasoning testMedium impact · October 13, 2026 · On October 13, 2026, Google posted a robotics note for Gemini 4 Argon. Google says pure reasoning scores 87 percent on Embodied Reasoning QA.

Key companies and projects

  • OpenAI
  • Google DeepMind
  • Anthropic
  • Meta

Related tags

  • open-weights
  • multimodal
  • embeddings
  • Mistral
  • coding
  • cybersecurity
  • Google DeepMind
  • Gemini
  • models
  • NVIDIA
  • speech
  • Ai2
  • evaluation
  • research
  • CUDA
  • Kubernetes
  • Liquid AI
  • edge
  • open weights
  • world model
  • preview
  • on device
  • Nemotron
  • OpenAI
  • computer use
  • pricing
  • Anthropic
  • markets
  • Moonshot
  • Kimi
  • Epoch
  • agents
  • Mirendil
  • pretraining
  • MiniMax
  • DeepSeek
  • forecasting
  • Reflection
  • Meta
  • Google
  • Microsoft
  • routing
  • ChatGPT
  • on-device
  • robotics

Name the layer

A product, an API model, and an open weight file can share a brand and still be different systems. GPT-6 Sol and GPT-6 Luna shipped in the same week under one numeral, aimed at different ChatGPT tiers, while Work and Codex stayed on their previous models. When a claim says “the model,” ask which surface, which size, and which date.

Three layers cover most of what we file. Frontier assistants are the systems people talk to. Open weights are the systems a developer can run and inspect. Specialist models do one job well: embeddings, gene finding, speech, a world model. A briefing that only covers the first layer is a product blog.

What we record

For a model release we want the date, the lab, the parameter count or the tier names if parameters are withheld, the license, and where it runs. We also want what did not move. A chat launch that leaves the coding agent unchanged is part of the story.

Benchmarks stay attributed. “Best in class” is the lab’s sentence until someone else measures it. We would rather print the suite name and the delta, as with EmbeddingGemma 2’s reported MTEB Code move from 68.76 to 78.68, than a medal.