Ollama 0.40 makes MLX the Mac default.
Ollama 0.40 is GA, and on Apple Silicon it now runs supported models on MLX by default
The release notes name the families that switch today: Qwen 3.8, Gemma 4, Qwen 3.6 and Qwen 3.5. The decision models Nimble, tev1, Clef and Clef Flash also run on MLX now, and the MLX engine now runs an embedding model too: Google’s new EmbeddingGemma 2. The last release candidate, also on Tuesday, added multimodal embeddings, so the embed endpoint accepts images alongside text. Ollama says it will keep testing and enabling more models.
Why this matters for Mac owners: until now the MLX engine was opt-in, release-candidate territory. From this release, an ordinary upgrade changes which engine runs your models. That is likely the faster path on Apple GPUs, but the notes give no speed figures, so anything you hear about speedups needs to come from your own measurements. The notes also say nothing about architectures MLX doesn’t support yet.
Before you upgrade a box you depend on: rerun the prompts you already benchmark, check memory use, and compare outputs, because a different engine can change more than speed. Keep the 0.35.1 installer around if you need repeatable results this week.
Sources: Ollama v0.40.0 release (Oct 6) · v0.40.0-rc6 (multimodal embeddings) · Ollama library: embeddinggemma-2
| Model / package | Org | Notes |
|---|---|---|
| Ollama v0.40.0 Oct 6 · GA |
Ollama | GA MLX default on Apple Silicon · decision + embedding models on MLX |
| EmbeddingGemma 2 Oct 6 |
Google DeepMind | Open 740M · text/image/audio/video · Apache 2.0 |
| Decisions API · gpt-6-luna Oct 6 · beta |
OpenAI | Cloud typed answers · $0.10/1M input · GA “coming weeks” |
| Mistral Large 4 Oct 6 · preview |
Mistral | Watch 1T / 49B active · API now, weights end of month |
| LiteRT-LM 0.18 / MediaPipe 1.1 Oct 6 |
Google AI Edge | On-device embeddings server · retriever · decision task |
| Transformers 5.19.0 Oct 6 |
Hugging Face | GA EmbeddingGemma 2 · MoE router-logit breaking change |
| TRL 1.14.2 Oct 6 |
Hugging Face | GA fixes silent GRPO/RLOO end-of-turn bug |
| Nano Banana 2.1 Oct 6 · GA |
Cloud Gemini API image model · old flash-image deprecated | |
| mlx-vlm main Oct 6 · unreleased |
Blaizzy / mlx-vlm | Watch EmbeddingGemma 2 + Kolibri merged · latest release 0.7.6 |
| llama.cpp tips Oct 6–7 · not GA |
ggml | Watch Metal quantized-FA fix · pplx-decider · K2 Horizon · GA still v0.6.0 |
| Item | Beat | Status |
|---|---|---|
| Anthropic CVP tiers Oct 6 |
Defense / Red Team / Specialized cyber access · data retention required | Active |
| Mistral Large 4 cyber preview Oct 6 |
Reduced moderation for vetted partners until weights ship | Watch |
| vLLM GHSA-6cxc Low · Oct 6 |
Stale token allowlist leaks across requests · fixed in 0.31.0 | Active |
| OpenAI usage tiers Oct 6 |
Five tiers become Build / Launch / Grow | Active |
| Claude Models API thinking field Oct 5 · correction |
Reports whether thinking can be disabled per model | Active |
| Sonnet 4.5 retirement Claude API · Nov 30 |
Migrate to Sonnet 5.5 · partner clouds set own schedules | Watch |
| Enterprise Frontier Safeguards Anthropic · “later this fall” |
Retention-free option for CVP members | Watch |
Critical · TRL upgrade · Ollama engine swap
If you ran GRPO, RLOO, or distillation on a Gemma 3/4 or Phi-3.5 model with TRL’s default generation path, upgrade to 1.14.2 and rerun. On Macs, Ollama 0.40 changes the engine under supported models, so re-benchmark before you rely on it. A fourth vLLM advisory (Low) is also fixed in 0.31.0.
Watch · ML4 weights · Decisions GA · edge rollout
Mistral says Large 4 weights arrive by the end of October, and the license is still unstated. OpenAI expects the Decisions API to reach GA in the coming weeks. EmbeddingGemma 2 in Android ML Kit and in Google’s Model Garden is “coming.” mlx-vlm has EmbeddingGemma 2 and Kolibri on main, but neither is in a release yet. OpenAI’s Ironclad post reports computer-use gains on its own eval.
Opportunity · simpler tiers · thinking flag · local embeddings server
OpenAI’s three usage tiers start at $5, $100, and $500 in total credit purchases. Claude’s Models API now tells you per model whether thinking can be turned off (Oct 5; Issue 32 missed it). LiteRT-LM’s serve exposes an OpenAI-compatible embeddings endpoint for local multimodal RAG.
Re-benchmark your Studio on Ollama 0.40
Before you upgrade, record the eval rate for two models you use daily with ollama run <model> --verbose on 0.35.1. Upgrade, then run the same prompts again on a model from the MLX list, such as Qwen 3.8 or Gemma 4. Compare speed, memory, and a few outputs side by side.
Build a local photo-and-notes search with EmbeddingGemma 2
Pull embeddinggemma-2 in Ollama and embed a folder of notes and screenshots through the embed endpoint. Use the card’s task prefixes for queries and documents. Store vectors at 768 and at 256 dimensions, and compare retrieval on 20 queries you know the answers to.
Race a local decision model against the hosted one
Label 50 real inputs, such as support tickets or routing calls. Run them through Clef on Ollama’s MLX engine and through OpenAI’s /v1/decisions with the same questions. Compare accuracy at your threshold, latency, and cost. Keep the inputs non-sensitive, or use your zero-data-retention settings.
Issue 33 · Wednesday, October 7, 2026 · Not investment advice · Sources cited inline · Window: after Issue 32 (Tue Oct 6 morning ET) through Wed Oct 7 morning ET · Issue 32 llama.cpp v0.6.0 / vLLM advisories / watermarks / GLM 5.3 not re-led · Ollama 0.40 RCs were Issue 31–32 watch items; GA is the new delta · llama.cpp tips after v0.6.0 are not GA · mlx-vlm main is unreleased · Mistral Large 4 weights not yet released · Correction: Claude release notes have an Oct 5 entry Issue 32 missed · Listen omitted (episode not published)
Tuesday morning after Issue 32 through Wednesday morning (Oct 6–7 ET). ASTGL / Fraser takes, sourced, not recap theater.
Your Mac’s default engine changed on Tuesday
Most people running Ollama on a Mac never opted into the MLX engine. They will get it anyway the next time they update, and I think that’s the right default. MLX was built for Apple silicon, and Ollama spent several release candidates getting decision models and embeddings onto it. But a default switch is not a free speedup. The release notes don’t include a single performance number and they don’t say what happens to architectures MLX can’t run yet. Treat 0.40 like a new engine, because it is one. Rerun your own prompts with timing turned on, check memory on the models you actually keep loaded, and diff a few outputs. If your work needs repeatable numbers this week, stay on 0.35.1 until you’ve done that.
Sources: Ollama v0.40.0 · v0.40.0-rc6
The classifier tier goes hosted
Lately the decision-model story has been a local one: Clef and Nimble quants, llama.cpp’s decision endpoint, mlx-vlm’s decision API, and SGLang’s own
/v1/decisionson Oct 2. On Tuesday OpenAI shipped a hosted version. The Decisions API is in beta ongpt-6-luna. You send text or images plus typed questions (yes/no, pick one, or score against a rubric) and get probabilities back. You pay $0.10 per million input tokens and nothing for output. OpenAI says it is about 10× faster than the Responses API. That is OpenAI-reported, so measure it. At that price, local decision models no longer win on cost. They win on privacy, offline use, and latency you control. The endpoint name matches SGLang’s, but I haven’t checked whether the request formats match, so don’t assume you can swap one for the other. Build a labeled set of 50 or 100 real inputs and run it through both before you choose.Sources: OpenAI API changelog (Oct 6) · Decisions guide
One embedder for everything, if you re-index
EmbeddingGemma 2 is the most useful local-AI release of the day. A 740M-parameter, Apache-licensed model puts text, code, images, audio, and video in one vector space. You can load only the text part (about 270M), and you can cut vectors down to 256 dimensions without losing much quality. It worked on day one in Ollama, llama.cpp, Transformers, and Google’s LiteRT, and MLX support is merged on main. For a Mac running local RAG, that’s an easy yes to try. Two cautions. First, new embeddings mean re-indexing everything you stored with the old model. Second, check the numbers yourself. Google’s model card says MRL cuts storage up to 6×, and its AI Edge post says up to 8×. Ollama’s library page lists a 256K context window, while Google’s card says 8K. Use the card’s figure until someone explains the difference. The benchmarks are Google-reported.
Sources: Google blog · Model card · Google AI Edge · Ollama library
Cyber capability is now a product tier
On the same day, two labs gave two different answers to the dual-use question. Anthropic folded Project Glasswing into an expanded Cyber Verification Program with three tiers. Defense Access covers SOC, incident response, and vulnerability validation. Red Team Access covers authorized pentesting. Specialized Access, for critical infrastructure, is reviewed with the US government. Mistral previewed Large 4 and pitched it as open weights for security work. Mistral claims closed models score near zero on its patch test because they refuse, and says the weights arrive at the end of the month. That claim is Mistral-reported. Meanwhile, ML4’s own pre-release cyber access goes to vetted partners and state authorities only. My read: refusal behavior is now a procurement criterion. Security teams should write down which tier or deployment they need before a deadline forces the question. Anthropic’s tier test results are Anthropic-reported, and ML4’s license isn’t stated yet.
Sources: Anthropic — CVP (Oct 6) · Mistral — Large 4 (Oct 6)
Check your clipped ratio
Two reminders on Tuesday that training numbers can be wrong without any error. TRL 1.14.2 fixes a bug where GRPO, RLOO, and distillation runs on Gemma 3/4, Phi-3.5, and similar models didn’t stop at the end of a turn, and then trained on whatever the model generated afterward. The only visible symptom was a completions clipped ratio near 1. Separately, Aleph Alpha wrote up the Kolibri decontamination work. When they put leaked benchmark documents back into a proxy run, HumanEval rose 45 points and MMLU rose 28. Their own detector caught only part of the leaks. If you fine-tuned a Gemma with TRL’s default generation path, upgrade and rerun. If a benchmark jumps after a data change, assume leakage until you can show it isn’t.
Sources: TRL v1.14.2 · Aleph Alpha (Oct 6)