Open-Weight Models Dominate October 2026 as Meta, Reflection Ship

A month that belonged to downloadable weights

The first half of October 2026 brought a lot of AI model releases, and the biggest share of the attention went to models whose weights anyone can download. Meta released Muse Spark Open. Reflection AI announced Beam. Mistral previewed a trillion-parameter Large 4. Smaller labs shipped specialized open models almost every day. The three large closed labs, Anthropic, OpenAI and Google, were still busy. Most of their October work, though, filled in cheaper tiers or widened access to models they had announced in September. They did not open a new frontier this month.

Release trackers show the same pattern. One running ledger counted eight models in the first two days of October, and four of them came with public weights under Apache 2.0 licenses: Cloudflare's Clef and Clef-flash, Amazon's Strands Decider 2B and Bilibili's Index-Translate-35B preview.27 Another tracker covered the seven days ending October 17 and listed seven releases. Four were open-weight: DeepSeek-V4.1-Flash-1017, Muse Spark Open and two NVIDIA Nemotron 3.5 models.30 One roundup said open weights "dominated" the first two weeks, starting with Aleph Alpha's Kolibri and building up to Meta's release on October 15.24

Meta's play: the "most capable US open-weight model"

The biggest launch of the month was Muse Spark Open. Meta's AI chief Alexandr Wang called it the most capable US open-weight model to date, and Meta says it competes with, and sometimes beats, Chinese models such as Qwen, Kimi K3 and GLM.21 The model card lists 816 billion total parameters, of which 42 billion are active per token. Context runs to 1 million tokens at inference, and the model accepts text, image, video and audio while producing text only.29 Meta pitches it for coding and agentic work and has paired it with its Muse Code terminal.22

The most unusual part of the launch is that Meta worked with Apple on it. Meta says the model is optimized for M5 Ultra chips and can run on two linked Mac Studios.21 That still takes a lot of hardware. TechCrunch, citing Engadget, says the model needs at least 1 terabyte of RAM.22 One ledger puts the BF16 weights at roughly 1.6 TB and an NVFP4 build at about 530 GB.27

The license comes with conditions. Products built on the model must carry Meta attribution, and companies with more than $20 million in revenue that want to offer it as a hosted service need a separate agreement.27 One guide's advice for commercial users is to check Meta's terms before shipping.25 So Meta's model is open in a narrower sense than the Apache 2.0 releases that make up much of the month's count.

Meta also shipped a closed model, Muse Wonder, a few days earlier and said it rivals the most advanced ChatGPT and Claude models.2130 Meta is now competing on both fronts, with a closed frontier model and a large open one released in the same week.

The Western answer to Chinese open models

The US open-weight launches share one stated goal: catching up with Chinese labs. The Guardian says Chinese developers have been the leading open-weight players for the past couple of years.21 TechCrunch described Reflection's Beam as a test of whether the West can build its own answer to DeepSeek, Qwen and Z.ai.23

Beam is a 501-billion-parameter mixture-of-experts model with 23 billion active parameters and a 1-million-token context window. It handles text only.23 Reflection says Beam matches Z.ai's GLM-5.2 on advanced reasoning while using three to four times less inference compute. TechCrunch points out that those claims have not been independently verified.23 Reflection admits that Kimi K3 is still ahead on raw capability, and it reports 80.9% on SWE-bench Verified.26 The company is promising Apache 2.0 weights this month.2526 Until those weights are actually published, Beam is an announced open model, not one developers can download.

Mistral's Large 4 fits the same pattern. Its documentation lists 1.05 trillion total parameters and 49 billion active, and the model is labeled a public preview.28 The weights are promised by the end of October, but the license is still undecided.25 That uncertainty limits what businesses can plan around it. Artificial Analysis gives it a score of 38, the same as OpenAI's GPT-6 Luna, but puts its cost at about $1.13 per task against roughly 7 cents for Luna.26 By that measure, one of Europe's biggest open models currently costs far more per task than OpenAI's cheapest tier.

The smaller releases may matter more in practice. Aleph Alpha's Kolibri-1 is an English-German mixture-of-experts model with 78.1 billion total parameters and about 3.46 billion active.24 Liquid AI's LFM2.5-24B-A2B activates just 2 billion of its 24 billion parameters and is aimed at on-device agents.27 JetBrains released Mellum 2.1, a 12B mixture-of-experts model for coding agents.28 On October 15, MBZUAI's Institute of Foundation Models and Inception42 released Jais 3, a family of six open Arabic models ranging from 900 million to 375 billion parameters.24 Beam's "AI factories" pitch and Jais 3 both show that sovereign deployment is becoming a reason in its own right to build open models.2324

Claude: completing the 5.5 family

Anthropic's October release was Claude Haiku 5.5 on October 7. It completed a staged rollout that began with Opus 5.5 on September 22 and Sonnet 5.5 on September 28.31 Anthropic says Haiku 5.5 costs about 75% less to run than Haiku 4.5. It is built for high-volume work such as summaries and classification and can act as a subagent under Opus and Sonnet in coding workflows.36 One tracker lists its price at 10 cents per million input tokens.26

The two September models set the direction. Opus 5.5 performs at about the level of Fable 5.1 on most work and costs 40% less to run than Opus 5.32 It is priced at $4 input and $20 output per million tokens, down from $5 and $25.38 It was also Anthropic's first release since CEO Dario Amodei called for slowing the frontier.38 Reports differed on Sonnet 5.5's price. One early piece said the API price had not been disclosed,40 while later timelines list $2 input and $10 output per million tokens.3337 Anthropic's own claims for Sonnet 5.5 are that it is 30%-plus faster and up to 30% cheaper for most work.34

GPT: cheaper Sol, and GPT-6 for everyone

OpenAI's October news was mostly about distribution. On October 7 it began rolling out GPT-6 with a feature called Intelligent UI in ChatGPT, starting with paid tiers. Free and Go users followed a day later.12 GPT-6.1 Sol is sometimes counted as an October release, but OpenAI introduced it at DevDay on September 29.1114 OpenAI says it comes close to GPT-6 Astra on agentic coding and computer use at one-fifth of Astra's token prices.13 On October 6, OpenAI also put its Decisions API into public beta, running on GPT-6 Luna.24

TechCrunch reported that OpenAI did not release a GPT-6.1 Astra. Citing the Wall Street Journal, it said the release was scrapped over safety concerns that came up in internal testing.14 OpenAI's pricing moves are notable too. Sol and Luna came in at half the price of the GPT-5.6 tiers they replaced.15

Gemini: Google returns to the frontier

Google made the month's biggest closed-model move. Gemini 4 Argon reached general availability on October 13 after first going to trusted cyber defenders through Google's Fairwind Program on September 30.1 It launched at an introductory price of $2 input and $10 output per million tokens.3 That matches GPT-6.1 Sol's price exactly,13 which suggests the main closed labs have settled on a common price for their workhorse models. Business Insider reported that the launch followed Google's decision not to ship the Gemini 3.5 Pro it had promised.5 One report adds that Argon scored 53 on Artificial Analysis's index. That beats Gemini 3.1 but still trails the top Claude and ChatGPT models.8

The reading

The headline count and the frontier tell different stories. By number of releases, October belongs to open weights. At the top end, the important closed models came out in late September, and October's closed releases mostly made them cheaper and more widely available. Prices across the closed labs are converging, with Argon and Sol both at $2/$10 and Opus cut to $4/$20.31338 Against that backdrop, the open-weight models' best argument is control over deployment, not the lowest cost per task. Several of the biggest open launches also come with caveats: weights not yet released, licenses still undecided, or terabyte-scale memory requirements. Whether that argument holds up will be clearer once Beam and Mistral Large 4 actually publish their weights.