Open-weight models, October 2026: GLM 5.3, Large 4, Kimi K3.1

October 2026 added six open-weight or open-weight-promised models worth knowing: Z.ai's GLM 5.3, generally available on Amazon Bedrock since 5 October, and GLM 5.3 Flash, on Mistral's platform since 22 October at $0.15 input, $0.03 cached and $0.50 output per million tokens; Mistral Large 4, in public preview since 6 October with weights promised by the end of the month; DeepSeek V4.1-Flash-1017, released with weights on 17 October; Moonshot's Kimi K3.1 Pro, with weights promised by 4 November; and Alibaba's Qwen-Image 2.1 image models, whose published weights are licensed for non-commercial use only. Whether any of them is good enough for a workload is a question for your own evaluation, run at the real cost per task on the provider you will actually use.

The background on licences and families is in open weight models in 2026, and the Chinese families in particular in the Chinese open-weight models guide. This article covers what changed in October.

The October releases at a glance

Each row in detail, with what its vendor actually says.

GLM 5.3. AWS made Z.ai's GLM 5.3 generally available on Amazon Bedrock on 5 October through US and Global cross-Region inference profiles. AWS describes a mixture-of-experts model with 753B total parameters and about 40B active per token, a 1M-token context window, up to 128K output tokens, reasoning that is always on with selectable effort, and explicit prompt caching; the post gives no price¹. Z.ai's own API lists $1.40 input, $0.26 cached input and $4.40 output². The weights are on Hugging Face under the GLM-5.3 License³.

GLM 5.3 Flash. Mistral added zai-glm-5-3-flash on 22 October: a multimodal mixture-of-experts model with 320B total parameters, 18B active and a 1M-token context, at $0.15 input, $0.03 cached and $0.50 output⁴. Z.ai lists the same three rates on its own API², and the weights are on Hugging Face under MIT⁵. Mistral also retires GLM 5.2 on its platform on 31 October in favour of GLM 5.3 at the same price⁴.

Mistral Large 4. Released in public preview on 6 October with launch pricing of "50% off for 2 weeks"⁴. Mistral's model page describes an open-weight, multimodal mixture-of-experts model with 52B active and 1.05T total parameters plus a 1.6B vision encoder, a 1M-token context, text and image input, and lists $1.36 input, $0.14 cached and $4.18 output⁶. The weights link reads "Coming soon"⁶; the announcement says "We will release the weights by the end of the month"⁷. No licence is named yet.

DeepSeek V4.1-Flash-1017. Released on 17 October with weights on Hugging Face and ModelScope; DeepSeek's notice does not state the licence⁸, and the Hugging Face model card lists MIT⁹. On DeepSeek's API, deepseek-flash is now served by this snapshot, with a 1M-token context, 384K maximum output, and $0.30 cache-miss input, $0.006 cache-hit input and $1.20 output at peak, half that off-peak¹⁰. Alibaba Model Studio added deepseek-v4.1-flash-1017 the same day¹¹.

Kimi K3.1 Pro. Moonshot's guide describes a 1M-token context, reasoning levels low, high and max with max the default, thinking that cannot be turned off, and says "the full model weights will be released by Nov 4"; no licence is stated¹². Moonshot's pricing page lists kimi-k3.1-pro at $2 input, $10 output and $0.20 cached input, and a Highspeed variant at $4 and $20¹³. Its predecessor, Kimi K3, is published under the Kimi K3 License¹⁴.

Qwen-Image 2.1. Alibaba Model Studio added qwen-image-2.1-pro on 2 October and qwen-image-2.1-turbo on 9 October, both in its International scope¹¹. The Hugging Face weights for Qwen-Image-2.1 and its Turbo variant are under the Qwen Research License¹⁵.

Licences: read the file, not the label

"Open-weight" covers very different rights. From the licence files the vendors publish:

The sources are Z.ai's GLM-5.3 licence³, the Kimi K3 licence¹⁴ and the Qwen Research License, which states the grant is "FOR NON-COMMERCIAL PURPOSES ONLY" and directs commercial users to request a licence¹⁵. These licences govern the weights. When you call a model through a provider's API, that provider's service terms apply as well, so a hosted Qwen-Image call and a self-hosted one are different legal positions. Have counsel read the actual file before a self-hosting decision; this table is a map, not advice.

Open weights do not mean the lowest list rate

It is tempting to assume an open-weight model is the low-cost option. Compare the hosted rates above with closed models' list rates read the same day:

The closed rates come from OpenAI¹⁶, Google, whose Gemini 3.8 Flash rate doubles on 1 January 2027¹⁷, and Anthropic¹⁸. The point is not that one column wins. It is that the per-token rate does not decide the question: a model with a lower list rate that needs more retries, longer reasoning or a larger prompt to pass the same task can cost more per completed task.

When an open-weight model is good enough

Open weights earn their place when one of these is true, and the frontier models keep theirs when it is not:

• You need control over where the model runs. Self-hosting, a region your provider does not offer, or an air-gapped environment. This is the reason no closed API can answer, and the reason licences matter.
• You need to adapt the weights. Fine-tuning or distillation into a smaller model you own.
• The task is bounded and checkable. Extraction, classification, routing, summarising to a template, code edits with tests. These are where a lower-rate model most often passes, and where a failure is visible.
• Volume makes the rate dominate. At high volume, the difference between $0.50 and $10 per million output tokens is the budget.

Stay with a frontier model where failures are expensive and hard to detect: long multi-step agent runs, ambiguous reasoning, and tool use where one wrong call is costly. And remember that two of the six "open-weight" models here have no downloadable weights yet; planning a self-hosted deployment on a promise is a risk to name in the plan.

An evaluation plan for a candidate

1. Freeze a test set from real traffic. A few hundred recorded requests per workload, with a check for each: an exact answer, a schema, a test suite, or a rubric a second model or a person applies.
2. Fix the parameters. Note that GLM 5.3 and Kimi K3.1 Pro reason by default with thinking always on¹¹²; set the effort explicitly so you are comparing like with like.
3. Measure run-to-run noise first. Run the incumbent twice on the same set. A candidate's difference smaller than that noise is not a difference.
4. Score cost per completed task. Input, output and reasoning tokens, retries and failures, at the rate you will actually pay on the provider you will use, including peak and off-peak rates such as DeepSeek's¹⁰.
5. Repeat on each host you plan to use. The same weights on Bedrock, Mistral and the vendor's own API are separate systems; the method for comparing hosts is in same open-weight model, different hosts.
6. Shadow, then canary. Send a copy of live traffic to the candidate without using its answers, then a small weighted share, and watch error rates and answer length before moving more.

Running open-weight models through INFRO

INFRO is a control plane in front of provider accounts you already hold; each provider bills you directly, and INFRO's charge is a separate subscription. For the models above, what it does today:

• Connections to the providers that serve them. Z.ai, Mistral (which INFRO also registers for Z.ai's models), DeepSeek, Moonshot, Alibaba Cloud Model Studio and Amazon Bedrock are connectable with your own keys, and an OpenAI-compatible endpoint connection reaches a self-hosted server such as vLLM.
• Catalog rates. INFRO's model catalog carries list rates for GLM 5.3, GLM 5.3 Flash, DeepSeek Flash and Kimi K3 as reference data for cost estimates. As of 27 October it does not carry Mistral Large 4, Kimi K3.1 Pro or Qwen-Image 2.1. On providers that take their own model ids, such as Mistral and Alibaba, a request can name mistral/<model id> for a model outside the catalog; it is served, but unpriced.
• Comparing hosts. Where two of your connections serve the same catalog model, a route can order or weight them, and usage grouped by provider and model shows errors, latency and estimated cost side by side. Traffic changes flag a connection that starts answering with a different model id or longer answers.

What INFRO does not do: it does not host or serve weights, it does not make two providers' copies of a model behave the same, and it does not judge answer quality; the evaluation above is yours to run. It also does not interpret licences.