The week's open models caught up

Week 43 of 2026 in AI models, for people who don't read model cards.

Most weeks, the news about AI models sounds like a list of names and numbers. This week was louder than usual, and the numbers hide a pattern: models you can download and run yourself got big enough to be taken seriously. Below is what happened, what the words mean, and what I would do with it. I also ran one prompt through seven models in FLORA, so you can see how different they are.

The short version

Meta, Reflection and DeepSeek all released open-weight models. Meta's is an 816-billion-parameter model, and it comes with a licence that has conditions.

Open weights is not open source. The difference matters the moment you build a product on one of them.

Google made Gemini 4 Argon generally available, MiniMax shipped M3.1 Flash, and OpenAI and xAI updated their voice models.

Kling 4 opened up, with clips of up to thirty seconds.

China proposed export controls on its own AI models and chips.

A quieter trend: decision models that answer with a probability instead of a paragraph.

1. Open models grow up

On 15 October Meta released Muse Spark Open. According to Meta's own developer blog, it has about 816 billion parameters, uses roughly 42 billion of them per token, takes text, image, video and audio as input, and has a context window of one million tokens. The weights are on Hugging Face. Meta calls it the most capable open-weight model from the United States. That is Meta's claim; the page gives no benchmark scores to back it up. TechCrunch covered the launch and so did the PyTorch blog.

A new lab, Reflection, shipped Beam: 501 billion parameters in total and about 23 billion active, text only. It was announced on 5 October and the weights followed on 16 October. Reflection says the licence is Apache 2.0, and that Beam matches leading Chinese open models on advanced reasoning at much lower compute cost. I could not find a page from Reflection itself, so the details come from TechCrunch and VentureBeat. Treat the performance claim as the company's own.

On 17 October DeepSeek updated V4.1 Flash to a version called 1017. Its changelog says the update improves performance while shortening tasks and outputs, and lists benchmark scores, for example 42.6 on Humanity's Last Exam without tools. You call it with the existing name deepseek-flash, and the API price stays the same. Coverage says the weights are MIT-licensed; the changelog entry itself does not state a licence.

In plain words: parameters, and why "active" matters

Think of a model's parameters as its dials. More dials give a model more room to store patterns. Muse Spark has 816 billion of them, but only about 42 billion work on any single word. That design is called mixture of experts: instead of one giant brain doing everything, the model has many specialists and a router that picks a few for each step. You get the knowledge of a very large model at something closer to the running cost of a much smaller one.

In plain words: a million tokens of context

A token is a small piece of text, roughly three quarters of a word. A context window of one million tokens means the model can hold about 750,000 words in view at once, which is roughly ten novels. It does not mean the model remembers all of it equally well. It means that much text fits in one conversation.

2. "Open" is a licence question

Open weights means you can download the trained model and run it yourself. Open source would also give you the code and the training data, with a licence that has no strings. Muse Spark Open is the first kind. Meta ships it under its own Muse Model License; coverage describes conditions such as a "Built with Muse Spark Open" notice in products and a separate agreement for large hosting companies above a revenue threshold. I did not read the licence text, and Meta's blog page does not summarise it, so read the file on Hugging Face before you build.

The rule of thumb I use: if a model matters for a product, check four things before you start. Can you download it? Can you run it yourself? Do you get the code and the data? And does the licence have conditions that touch your business?

3. The closed side

Google made Gemini 4 Argon generally available on 13 October. The Gemini API release notes confirm the date, and Google Cloud's documentation lists the model as GA. "Generally available" means anyone with access can use it, not only early testers. Coverage reports a limited-time launch price that doubles on 1 December; I could not confirm the numbers on a Google page, so check before you budget.

MiniMax released M3.1 Flash on 19 October. Its launch post claims inference that is nearly three times faster and a 70 percent lower cost for agent tasks, priced at $0.15 per million input tokens and $0.60 per million output tokens. It is available through the API, which is simply the way software talks to a model. The weights are promised for later.

4. Voice

On 15 October OpenAI began rolling out GPT-Live 1.1 in ChatGPT Voice, with new voices in six more languages plus a new English voice. The ChatGPT release notes are the primary source, though I could not open that page myself. Coverage counts thirteen new voices. API access has not been confirmed.

A day earlier xAI released Grok Voice 3.0 in its Voice Agent API at $0.08 per minute, the same price as the version before. xAI says the model ranks first on a speech-to-speech index; another test I found placed it second. xAI's own pages were not available to me, so the details come from Unite.AI.

Speech-to-speech just means a model that listens and answers in voice directly, without first turning your speech into text and the answer back into speech. That removes a step, and with it some delay.

5. Video

Kling 4 opened up in mid-October, with clips of up to thirty seconds from a single sentence, up from fifteen in the previous version. Kling's own page was not among what I could find, so this rests on coverage such as Kie.ai. Some features, like HDR, are listed as coming later.

6. A quiet trend: decision models

Most language models write text. Decision models do something else: you give them a situation and a question, and they answer with a probability. Is this support ticket urgent? Yes, at 87 percent. That is easier to plug into software than a paragraph, because code can act on a number.

Cloudflare released two of them on 1 October, Clef and Clef-flash, under Apache 2.0. According to Cloudflare's changelog they return probabilities for yes/no, choice and score questions, take up to 64 questions per request, and follow the System One API shape, so an existing integration can switch by changing the endpoint and model. I use one of these in my own work, Jev from TypeSafe, so I am not neutral here. Other companies shipped similar ideas this month, but I only verified Cloudflare's.

7. Outside the labs

On 16 October China's ministries proposed adding AI technologies to its export control list. Bloomberg and PBS report that the draft covers advanced models, agents, chip designs and computing clusters, with public comments open until 15 November. Coverage differs on the exact numbers, so read the details in the sources. Semafor adds the US angle.

Export controls are rules about what technology may leave a country. Until now the debate was mostly about the United States limiting chips going to China. This week the same idea pointed the other way.

Investors are also asking harder questions about how AI infrastructure is funded. I only saw a headline on that, so I am leaving it at that.

8. I ran one prompt through seven models

I wanted to show the difference between models instead of describing it. In FLORA I ran the same prompt through four image models and three video models. The prompt: a man with a short beard, a black bucket hat and round blue sunglasses, in an orange tactical vest, pointing at a big poster that says OPEN MODELS.

Model, Type, Price for this run & What I saw

FLUX 3 Image Klein, image, $0.009 - simple and cheap, a plain poster Qwen Image 3.0 Pro, image, $0.048 - clean flat illustration, clear text. Nano Banana 2.1, image, $0.052 - busy collage, the text gets mixed into the pattern. Grok Imagine Quality, image, $0.06 - stylised, the text runs over the sign. MiniMax H3 Max, video, 4 s, $0.38 - photorealistic studio, he walks up and points. Kling 3.0 Standard, video, 4 s, $0.44 - flat illustration, walks toward the camera. Kling v4, video, 4 s, $0.66 - illustrated style, ends in a close-up.

1 prompt and 1 run per model is not a benchmark. It tells you that the same words land very differently, and that price does not tell you which one fits your job.

What I would do with it

Pick the model for the job, not the model of the week. Test it with your own material. Check the licence before you build on it. And keep an option you can run yourself, because the open models just got good enough to be a real option.

I do the geeking, and you get the easy part. If you are building something and want to know which of these fits, send me what you have.

What I could not verify

I did not find Reflection's, xAI's or Kling's own announcements, so those sections rest on coverage.

I could not open OpenAI's release notes page, and the count of thirteen voices comes from coverage.

Google's pricing and the licence terms of Meta's model are not confirmed on pages I could read.

The claim about one million tokens equalling ten novels is a rough comparison.

The film's voice and music are AI-generated (@ElevenLabs v4 and ElevenLabs Music via @FLORA ), and so is the character animation. The text and research are mine, with @Anthropic Claude's help. All claims from companies are labelled as such, and sources are linked above.