October 2026 had two stories. At the top, Google made Gemini 4 Argon generally available and then announced Gemini 4 Carbon, which took first place on the Artificial Analysis Intelligence Index, but only for selected cyber defenders. Below the frontier, three Western labs released open weights at sizes that had been mostly Chinese: Reflection Beam at 501B, Meta's Muse Spark Open at 816B and Mistral Large 4 at about 1 trillion parameters. Moonshot's Kimi K3.1 Pro, at 2.8 trillion, kept China ahead on raw size.
The Releases at a Glance
| Date | Model | Lab | Type | Key spec |
|---|---|---|---|---|
| Oct 5 / Oct 16 | Beam | Reflection AI | Open weights (Apache 2.0) | 501B total, 23B active, text only |
| Oct 6 / Oct 29 | Mistral Large 4 | Mistral | API preview, then open weights | About 1T total, 49B active, 1M context, $0.70 / $2.10 |
| Oct 13 | Gemini 4 Argon (GA) | Closed | 1M context, $2 / $10 introductory | |
| Oct 15 | Muse Spark Open | Meta | Open weights (Muse Model License) | 816B total, 42B active, 1M context |
| Oct 23 | Kimi K3.1 Pro | Moonshot | Closed at launch, weights promised by Nov 4 | 2.8T total, about 104B active, 1M context |
| Oct 28 | Gemini 4 Carbon | Closed, limited access | AA Intelligence Index 59, $2 / $10 introductory | |
| Late Oct | Apriel Voice 27B Preview | ServiceNow | Open weights (MIT) | Speech in, text and tool calls out |
Why October 2026 Matters for Brand Visibility
Two shifts. First, the best model in the world is no longer one the public can use. Carbon reaches cyber defenders through Google's Fairwind Program first, and Argon took the same route in September before going GA. Leaderboard positions now say less about what consumers see. Second, huge open models went Western. Enterprises that wouldn't deploy Chinese weights now have three US and European options at 500B to 1T parameters, so more brand questions will be answered by models running inside companies, where nobody outside can see the answer.
Google: Gemini 4 Argon and Carbon
Argon, announced September 30 for trusted defenders, became generally available on October 13 at an introductory $2 / $10 per million tokens. Carbon followed on October 28. It scores 59 on the Artificial Analysis Intelligence Index v5.0, ahead of Claude Opus 5.5 at 58 and GPT-6 Astra at 57, and Google reports 88.2% on SEC-bench Pro. Google says developers and consumers will get Carbon as soon as possible, with no date. Argon is what powers Gemini products now, so test brand prompts there. See the Gemini 4 Carbon FAQ and Carbon vs Opus 5.5 vs Astra.
Mistral: Large 4 Open Weights
Mistral opened the API preview on October 6 and published weights on October 29 as Mistral-Large-4-1T-A52B, under a modified MIT license that restricts hosted-service resale. It has about 1T total parameters with 49B active, takes images, and has a 1M context. Mistral says its 26 on the Artificial Analysis index is the highest for a Western open-weight model. See running Mistral Large 4 locally.
Meta: Muse Spark Open
Released October 15 with 816B total and 42B active parameters and a 1M context, under Meta's own license. Apple's MLX team published Mac Studio speeds the same week: 15.7 to 36.7 tokens per second on two to four M5 Ultra machines at 4-bit. See Muse Spark Open hardware requirements.
Reflection AI: Beam
Announced October 5, with weights on Hugging Face on October 16 under Apache 2.0. It has 501B total parameters and 23B active, and is text only. Hosted prices start around $0.25 / $1.00 per million tokens. Reflection reports 80.9 on SWE-bench Verified. See the Reflection Beam brief.
Moonshot: Kimi K3.1 Pro
Released October 23 as a post-trained Kimi K3, with a faster Highspeed variant. Weights were promised by November 4. Secondary coverage gives $2 / $10 per million tokens. See the Kimi K3.1 Pro brief.
ServiceNow: Apriel Voice 27B Preview
An MIT-licensed speech model built on Gemma 4 26B-A4B. It listens and returns text and tool calls, but doesn't speak back. See the Apriel Voice brief.
What to Test This Month
1. Re-run your prompt set on Gemini with Argon and keep it as the baseline for Carbon's public release.
2. If you sell to enterprises, test your category on Mistral Large 4, Muse Spark Open and Beam through a hosted API. These are the models that may sit inside your buyers' internal tools.
3. Compare which sources each model cites for your category. Open models trained on different mixes may favour different review sites and forums.