MiniMax opens M3.1 Flash, and a Studio can just hold it.
MiniMax put M3.1 Flash’s weights on Hugging Face overnight, and by morning there was a 3-bit MLX build that fits a 256 GB Mac Studio
M3.1 Flash is a large mixture-of-experts model aimed at coding, agent work, computer use, and security tasks. It has about 428B parameters in total, about 23B active per token, takes text, images, and video, and MiniMax says it handles up to 1M tokens of context. MiniMax trained it with quantization built in, so the official files are already compressed, and they include a small draft model for faster decoding. MiniMax’s benchmark table compares it with frontier models on coding and agent tests. By MiniMax’s own count it still trails Claude Opus 5 on Terminal-Bench 4.0 (38.1 vs 51.8). Every number is MiniMax’s (vendor-reported), and the card warns that third-party quantized builds may score differently.
On a Mac: the MLX community’s 3-bit text build is about 187 GB and is the realistic fit for a 256 GB Studio. The 4-bit build keeps MiniMax’s original format but needs about 249 GB, which leaves almost nothing for macOS or context. Both drop the vision part for now. mlx-lm’s maintainers merged support for loading MiniMax’s original files early this morning, but it isn’t in a tagged release yet. Unsloth’s GGUF repository for llama.cpp was still empty at our check, and Ollama offers the model only in its cloud.
Why it matters: this is a frontier-scale agent model that one 256 GB Mac can hold, and MLX support arrived the same night the weights did. The 3-bit build leaves roughly 70 GB for macOS and context.
Cautions. The license is the MiniMax Community License, not an open-source license. Its opening grant covers non-commercial use. Commercial use, which explicitly includes hosted APIs and fine-tuned derivatives, requires a “Built with MiniMax M3.1-Flash” credit and a notice to MiniMax, or written approval if your product earns more than $20 million a year. Military use is banned. Ollama’s cloud listing says 976K context where the card says 1M, and one community build’s card overstates the license by saying all commercial use needs prior approval. The only Mac-relevant speed figure, 23.8 tokens per second on the full original checkpoint, comes from one test run in the mlx-lm pull request, with no hardware named.
Sources: Model card · License · mlx-lm #2003 · MLX 3-bit · MLX 4-bit · Ollama
OpenAI announced a safety partnership with Poland’s NASK research institute. OpenAI
Model & runtime releases
| Item | Date | Badge |
|---|---|---|
| MiniMax M3.1 Flash weights (428B / 23B active, MiniMax Community License) | Oct 22 | Open |
| mlx-lm #2003: load original MiniMax M3.1 checkpoint (merged; not yet released) | Oct 23 | Open |
| mlx-community MiniMax M3.1 Flash Text 3-bit (~187 GB) / 4-bit (~249 GB) + oQ builds | Oct 23 | Open |
| Ollama v0.40.9: Phonon speech, d3 decision models, Flint on Linux/Windows | Oct 22 | Voice |
| Gradium Phonon 31M English TTS (Apache 2.0); 40M variants CC BY-NC | Oct 19–22 | On-device |
| llama.cpp b11833–b11836: DFlash2 probabilistic drafts, Nemotron-3.5-Embed fix, MiniCPM-OCR | Oct 22–23 | Open |
| Tencent WeVisOne-VL 2B / 4B vision models (Apache 2.0) | Oct 22 | Open |
| AI2 OlmoEarth v1.3 Earth-observation embeddings | Oct 22 | Open |
| OpenAI Decisions API GA (GPT-6 Luna, 400 questions/request) | Oct 22 | Cloud |
| Claude Models on Google Cloud (Anthropic managed) | Oct 22 | Cloud |
| SageMaker AI training session APIs | Oct 22 | Cloud |
| AgentCore Gateway: Bedrock InvokeModel + model fallback | Oct 22 | Cloud |
| NVIDIA Minerva (AIOps for AI factories) | Oct 22 | Cloud |
Security / policy / access
| Item | Date | Badge |
|---|---|---|
| AWS Continuum for code vulnerabilities GA (sandbox-validated fixes as PRs) | Oct 22 | Active |
| Agent Skills donated to the Agentic AI Foundation (Linux Foundation) | Oct 22 | Active |
| MiniMax Community License: non-commercial grant; credit + notice for commercial use; approval above $20M revenue | Oct 22 | Watch |
| Google Agent Compute Engine (agent microVMs); preview “in the coming months” | Oct 22 | Watch |
| php2shell CVEs still not on CISA KEV (catalog 2026.10.21) | Oct 23 | Watch |
| No new vLLM advisories since the Oct 20 batch; none for llama.cpp / Ollama / MLX / SGLang | Oct 22–23 | Quiet |
Critical
Before you ship anything built on MiniMax M3.1 Flash, read its license. Commercial use, including a hosted API or a fine-tune, needs a visible credit and a notice to MiniMax. If you use Ollama’s non-English Phonon voices, check the license: they are non-commercial only.
Watch
MiniMax M3.1 Flash in a tagged mlx-lm release, with vision support; GGUF builds for llama.cpp (Unsloth’s repo was empty at our check); and a local Ollama tag. Google Agent Compute Engine’s preview date. Still pending: Nemotron 3.5 Bolt E4B (NVIDIA said Oct 23; not on Hugging Face at our check), PyTorch 2.15 final, Flint in mlx-lm, and OpenAI’s gateway approvals ahead of Nov 20.
Opportunity
Nobody has published MiniMax M3.1 Flash numbers on a named Mac. Tokens per second and real coding-agent results for the 3-bit MLX build on an M3 Ultra, set against MiniMax’s own table, would be the first.
Run MiniMax M3.1 Flash on a 256 GB Studio
Use the 3-bit text build, which peaks around 187 GB: pip install -U mlx-lm, then mlx_lm.generate --model mlx-community/MiniMax-M3.1-Flash-Text-3bit --temp 1.0 --top-p 0.95 -p "Explain this stack trace: …" -m 2048. The download is about 187 GB, so check your disk first. Close other big apps, and keep prompts short until you know how much context fits. Text only for now.
Make Ollama talk
Update to Ollama 0.40.9, then run ollama run phonon "Hello from Ollama.". The default phonon tag is Gradium’s 31M English model, licensed Apache 2.0, at about 104 MB. Apps can stream the same voice through the OpenAI-compatible /v1/audio/speech endpoint. Skip the de, fr, es, pt, and multilingual tags for anything commercial.
Try a local yes-or-no model
Ollama’s own example: ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?". Run a few dozen of your real support or triage messages through it and compare its calls against your own before you put it in front of an agent.
Wednesday morning through Friday morning (Oct 21–23 ET). ASTGL / Fraser takes, sourced.
The 256 GB line: 3-bit fits, native 4-bit doesn’t
We’ve had MiniMax M3.1 Flash on Watch for three issues. The weights finally showed up around midnight, and the MLX community had builds up before most of us were awake. The useful detail for Studio owners is simple. The 4-bit build keeps MiniMax’s original format, but at about 249 GB peak it leaves a 256 GB machine with no room for anything else. The 3-bit build, at about 187 GB, is the one to try. My take: download the 3-bit build this weekend and test it on agent tasks you already know the answers to, because every benchmark so far is MiniMax’s. And read the license before you build anything on it. “Open weights” is accurate. “Open source” isn’t. If this ends up in a paid product, you owe MiniMax a credit line and an email.
Sources: Model card · License · MLX 3-bit · MLX 4-bit
Picking your data processor is now a menu option
Claude has been on Google Cloud for a while, run by Google. As of Thursday you can also buy Claude through Google Cloud with Anthropic running it. The price is the same and the billing is Google’s either way. The difference is who processes your data and when you get new features. With Anthropic managing it, you get things like Managed Agents and skills on day one, under Anthropic’s terms. With Google managing it, you keep Google as processor, plus VPC Service Controls, customer-managed keys, and regional data residency, but new features arrive later. My take: this is a compliance decision, not a model decision, so put your security and legal people in the room before engineering picks one. One gotcha from Anthropic’s setup docs: the Anthropic-managed option creates a brand-new Anthropic organization, separate from any you already have. Existing discounts don’t apply to past usage, so talk to your account rep first.
Sources: Google Cloud (Oct 22) · Claude release notes (Oct 22) · Setup docs
Agent Skills goes neutral, and the security guidance is the part to watch
Anthropic handed Agent Skills, the SKILL.md folder format for teaching an agent a task, to the Linux Foundation’s Agentic AI Foundation. That’s the same home MCP went to. Anthropic says more than 70 agent platforms support the format, including ChatGPT, Cursor, GitHub Copilot, and VS Code, and that a million skills have been published (Anthropic’s figure). My take: neutral ownership helps, because a format controlled by one vendor is a risk to build on. But the line I care about is on the roadmap: security guidance for writing and installing skills. A skill is instructions plus scripts that an agent will follow and run. That makes installing a stranger’s skill a supply-chain decision, and today there is no standard way to vet one. Until that guidance exists, read every skill before you install it, the same way you’d read a shell script piped from the internet.
Sources: Anthropic (Oct 22) · agentskills on GitHub
Yes-or-no answers got their own cheap endpoint, in the cloud and on your Mac
On the same day, OpenAI made its Decisions API generally available and Ollama shipped d3, a local family of models built for the same kind of job. Both take a piece of text, data, or an image and return a bounded answer: yes or no, one choice from a list, or a score. OpenAI charges $0.10 per million input tokens with no output-token charge, and says it is up to 10 times faster than asking the same model through its general API (OpenAI’s claim). My take: much of what agents do all day is routing, triage, and gating, and none of that needs a paragraph of reasoning back. Pulling those calls into a purpose-built endpoint is cheaper and easier to test. If the data can’t leave your network, try d3 locally first. If it can, Decisions is a cheap way to run thousands of these.
Sources: OpenAI (Oct 22) · OpenAI changelog · Ollama v0.40.9
Fewer subagents, stronger sandboxes
Two posts on Thursday point the same way. Anthropic’s applied AI team says many enterprise agents improved after folding their routers, planners, and domain subagents back into one main loop. In their words, “one clear system prompt, 4-5 thin data tools, and firm output guardrails” covered most of the use case. Google, meanwhile, previewed Agent Compute Engine: lightweight virtual machines for agents with snapshot and restore and short-lived credentials. It isn’t available yet; Google says the preview comes “in the coming months.” My take: the multi-agent diagrams from last year are getting simpler, and the hard problems are moving to the edges, meaning what the agent is allowed to touch and how quickly you can undo it. If your agent has five subagents, try collapsing them into one loop with fewer tools, and compare quality and cost.
Sources: Claude blog (Oct 22) · Google Cloud (Oct 22)