Two open models, and no weights yet

Two companies spent the start of this week claiming the same title. On Monday Reflection AI, a Brooklyn start-up founded in 2024 by two former Google DeepMind researchers, introduced Beam, which it says "advances the Western open-weight frontier". On Tuesday Mistral, the Paris lab, launched a public preview of Mistral Large 4, which it says significantly outperforms "any open-weight model developed in the US or Europe". Both pitch themselves as the answer to the Chinese open models that have led the category. Neither model can yet be downloaded.

That is the detail the launch coverage tends to pass over. Mistral says it "will release the weights by the end of the month"; TNW reported the date as 27 October. Until then the public can use the model only through Mistral's API. Reflection says Beam is "undergoing final red-teaming and evaluations", is open for now to "a select group of users" via a waitlist, and that the weights will follow "later this month" under an Apache 2.0 licence. Artificial Analysis, the independent benchmarking firm, still lists Mistral Large 4 as a proprietary model because its weights have not been released, The Decoder reports.

The interval is not idle, and Mistral's description of it is the most revealing sentence in either announcement. Until the weights ship, it says, "we are red-teaming the model in real-world settings with cybersecurity leaders, vetted partners, and state authorities, who will access the same model with reduced moderation and expanded cyber capabilities." Pierre Stock, Mistral's vice-president of science, told TechCrunch: "In the meantime, we'll work with trusted partners and governments to make sure that the open source weights can be used to defend, but not to [perform] malicious attacks." For three weeks, in other words, the least restricted version of an open model is available only to a chosen few.

Cyber defence is the centre of Mistral's case, and by Mistral's own account its top score owes something to other companies' policies. On one test in the Artificial Analysis Cyber Index, which asks a model to reproduce a real vulnerability in open-source software and then patch it, Mistral says Large 4 scores 82 per cent, the highest of any model. "Several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on the same test because they refuse to perform the task," the company writes. The Decoder's verdict: "So the test measures provider policies as much as model capabilities." Mistral turns that into its pitch. "The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks," Guillaume Lample, its co-founder and chief scientist, said, according to TNW.

The same announcement also claims the opposite virtue. Mistral says Large 4's average refusal rate on malicious cyber prompts drawn from JailbreakBench, StrongREJECT and AgentHarm "is higher than all OSS models." It does not explain how the model separates a defender reproducing a flaw from an attacker doing the same thing; The Decoder noted the omission. The practical point is about who sets that line after release. Once the weights are public, TNW notes, "a bank or a government can download the model, run it on its own machines, and use it to protect its own networks without asking anyone's permission." Mistral's moderation applies to its own API. On a customer's hardware, the customer decides how the model is run.

That is also the sovereignty argument Mistral has been making for months. In May, its chief executive Arthur Mensch told a French parliamentary commission that the French military's codebases should not be scanned by Anthropic's Mythos, The Decoder reports. Mistral says Large 4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres, and that the preview runs on the same machines; Stock gave TechCrunch a figure of 4,000 GPUs, "which is two to three times less than our Chinese competitors, and significantly less than the closed source competitors." Christine Lagarde has made a related point, having cited the summer's suspension of two Anthropic models after a US export-control directive as a warning for European finance.

On capability, independent measures put both launches some way short of the top. Artificial Analysis's Intelligence Index gives Large 4 38 points, up from 9 for Mistral Large 3, and ahead of Z.ai's GLM-5.2, but behind several Chinese open models and well behind Claude Opus 5.5 (Max) at 58, according to The Decoder. Reflection's own table is candid about its position: on DeepSWE, an agentic coding test, Beam scores 44.4 against 61.0 for GLM-5.3, 68.0 for Kimi K3 and 74.2 for DeepSeek V4.1 Flash. Mistral's chart, as reported by TNW, puts Large 4 at 62 per cent on the same test and Beam at 44. Reflection's argument is efficiency: it says Beam matches GLM-5.2 on advanced reasoning while using three to four times less inference compute. TechCrunch noted that its performance claims have not been independently verified.

Both companies are well funded for a second attempt. Mistral calls Large 4 the first milestone on a roadmap funded by its €3bn Series D, which it calls the largest equity round raised by a European technology company, and that its reinforcement-learning run "is still in flight". Reflection has raised roughly $4.7bn, according to PitchBook data cited by TechCrunch, and has signed compute deals with SpaceX and Nebius worth more than $7bn. It says Beam's RL phase used 10,500 Nvidia GB300 GPUs for four weeks. Mistral is charging $0.68 per million input tokens and $2.09 per million output tokens during the preview, The Decoder reports, while its documentation lists double those rates.

The test for both comes later this month, and it is a simple one. An open-weight model is one whose weights anyone can download. Until 27 October for Mistral, the date TNW reported, and an unspecified date for Reflection, these are products with a promise attached, and the least restricted version of Mistral's is in the hands of partners and governments it has chosen.