Two Open-Weight AI Models Launched a Day Apart: What Mistral Large 4 and Reflection’s Beam Actually Offer

On October 5, 2026, the US startup Reflection AI previewed Beam, its first open-weight model. The next day, October 6, French lab Mistral previewed Mistral Large 4, nicknamed “Le Chonk.” Neither model’s weights are downloadable yet — both are API-only previews, with full weight releases promised before the end of October 2026 — but together they’re the clearest sign yet that “open-weight” AI (models whose parameters you can download and run yourself, instead of only calling through someone else’s API) is becoming a real contest outside China, not just a side category led by DeepSeek, Qwen, and Z.ai’s GLM.

What “open-weight” actually means, and why “sovereignty” keeps coming up

Most AI products you use day to day — ChatGPT, Claude.ai, Gemini — only let you send a model a prompt and get a response back; the underlying parameters (the model’s actual “weights”) stay on the provider’s servers. An open-weight model publishes those parameter files so anyone can download them and run the model on their own hardware, whether that’s a laptop, a company’s private cloud, or a government data center. You still can’t see the training code or the training data in most cases — that’s “open source” in the fuller sense, and it’s rarer — but you get to control where the model runs and who can see what goes in and out of it.

That control is what both companies are selling as “sovereignty”: the ability for a company, hospital, bank, or government to run a capable model entirely inside its own walls, under its own country’s law, without a foreign cloud provider ever seeing the data. It’s become a bigger pitch in Europe specifically because of GDPR and the EU AI Act, and because several governments have made public noises about not wanting their critical AI infrastructure to depend entirely on US clouds.

What Mistral launched

Mistral’s own announcement describes Large 4 as a natively multimodal (text and image) model with a granular mixture-of-experts architecture that blends instruction-following and step-by-step reasoning in one model, about 1 trillion total parameters with 49 billion active per token. Mistral’s own technical documentation page for the model gives a slightly different figure — 52 billion active parameters and 1.05 trillion total, plus a separate 1.6 billion-parameter vision component — and as of this writing Mistral hasn’t reconciled the two numbers in one place, so treat “about a trillion parameters, call it 50 billion active” as the honest summary rather than a single precise figure.

The model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs inside Mistral’s own European data centers, according to the company, and the public preview runs on that same infrastructure. Mistral’s documentation lists a 1 million token context window for the preview.

How it benchmarks, with the caveats that matter

All of the numbers below come from Mistral’s own announcement, so treat them the way you’d treat any vendor’s self-reported scores: directionally informative, not independently verified. On Mistral’s figures, Large 4 scores 61.7% on the DeepSWE v1.1 coding benchmark and 59.9% on AutomationBench (a suite of 657 business workflows across apps like Gmail, Sheets, and Slack), putting it, by Mistral’s own account, ahead of DeepSeek V4 Pro and Alibaba’s Qwen3.8 Max on several of those tasks. A third-party aggregator’s own cross-check of published numbers found the opposite on one specific benchmark, DeepSWE v1.1: Kimi K3 and DeepSeek V4.1 Flash both post higher published scores there than Large 4’s 61.7%. Different open models lead on different tests, which is normal and part of why single-benchmark bragging rights are a weak signal on their own.

Mistral’s most specific claim is on cybersecurity: on an independent benchmark from Artificial Analysis that asks a model to reproduce a real software vulnerability and then patch it, Mistral says Large 4 scores 82%, the highest of any model tested there, while Claude Opus 5.5 and GPT-6 Astra reportedly score near zero on the same test — not because they can’t do the work, but because their safety filters refuse to attempt it. Mistral also ran a blind human evaluation with Surge AI on coding output quality, where professional annotators rated Large 4 second of five models tested (ahead of Kimi K3 and two GLM versions, behind Claude Opus 5).

A day earlier: Reflection AI’s Beam

Reflection AI, a US startup founded by former Google DeepMind researchers, previewed Beam on October 5. It’s a smaller, text-only model: 501 billion total parameters with 23 billion active per token, pretrained on 23.8 trillion tokens, with reinforcement learning run across more than 100 million practice rollouts on roughly 10,500 Nvidia GB300 GPUs over about four weeks, according to Reflection’s own announcement. Unlike Mistral, Reflection has already committed to a specific open-source license: Apache 2.0, with weights, a technical report, and developer tooling due “later this month.”

Reflection’s pitch is efficiency rather than raw size. The company says Beam matches Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute, and that the gap is even larger against bigger models like Qwen3.8 Max. On raw capability, Reflection’s own numbers show Beam trailing frontier open models: on Terminal-Bench v2.1, Beam scores 80.1 against Kimi K3’s 88.3 and DeepSeek V4.1 Flash’s 90.6, and on DeepSWE v1.1 it scores 44.4 against those models’ 68.0 and 74.2. Reflection frames this as an intentional trade-off — a “workhorse” model that’s cheaper to run per answer, not a leaderboard-topper — and notes its compute-efficiency comparison is its own estimate, built from assumptions about generation cost, not a measured, independently audited number.

Why the timing isn’t a coincidence

Both launches follow big recent funding rounds built on similar arguments. Mistral closed a €3 billion ($3.5 billion) Series D on September 8, 2026, led by Samsung Electronics with EQT’s Scaleup Europe Fund and PSG Equity as co-leads, pushing its valuation past €21 billion (about $24 billion) — nearly double its December 2025 valuation, and what Mistral and multiple outlets, including TechCrunch, described as the largest equity round ever raised by a European technology company. CEO Arthur Mensch tied the raise directly to this kind of release, saying it would let Mistral “move faster on R&D, products, infrastructure” for organizations that want to “industrialise AI inside systems that are already complex: their own data, their own tools, their own compliance constraints.” Reflection, for its part, has raised roughly $4.7 billion since 2024 from backers including Nvidia and Sequoia, and is explicit that Beam exists to give Western developers and governments an alternative to relying on Chinese open models for self-hosted AI.

Neither company has published independent revenue figures, and some financial commentary on Mistral’s round (including from the newsletter Value Add VC) has pointed out that a valuation built on “sovereignty” rather than disclosed revenue or best-in-class benchmarks is a bet on policy and government procurement preferences holding steady, not a guarantee that the underlying model is the most capable one available.

What this actually means if you’re not a developer

If you’re an everyday chatbot user, neither release changes what’s in your hand today — Mistral Large 4 isn’t in Le Chat yet, and Beam isn’t in any consumer product. Where this matters more directly: if you work somewhere that’s evaluating AI vendors for regulatory or data-residency reasons (healthcare, finance, government, EU-based companies generally), “can we run this ourselves, under our own law” is becoming a real item on vendor comparison sheets, and these two releases are what that conversation is increasingly about. If you’re just AI-curious, the practical takeaway is simpler: the gap between “open” and “closed” frontier models keeps narrowing, and that competition is part of why API prices for capable models keep falling.

FAQ

Can I actually download and run either model right now?

No. As of this writing (October 7, 2026), both are preview-only: Mistral Large 4 is reachable through Mistral’s API/Mistral Studio, and Beam is available to a limited early-access waitlist through Reflection’s platform. Both companies say downloadable weights are coming by the end of October 2026.

Is “open-weight” the same as “open-source”?

Not quite. Open-weight means the trained parameters are published for anyone to download and run. Open-source, in the fuller sense, usually also means the training code, training data, or both are available too. Reflection has said it will also open-source some of its RL training tools and safety-evaluation methods alongside Beam’s weights; Mistral’s announcement doesn’t make that broader commitment for Large 4.

Which one is “better”?

They’re not really aimed at the same job. Large 4 is bigger, multimodal, and pitched at frontier-level capability across coding, cybersecurity, and knowledge work; Beam is smaller, text-only, and pitched at being cheap to run per answer while staying competitive with mid-tier Chinese open models. Both sets of benchmark numbers are self-reported by the company that built the model, so independent, apples-to-apples testing once both ship weights will tell you more than either company’s own charts.

Why does it matter that these are European and American companies specifically?

Open-weight AI has been led for the past couple of years largely by Chinese labs — DeepSeek, Alibaba’s Qwen, Z.ai’s GLM, Moonshot’s Kimi. Both Mistral and Reflection explicitly frame their new models as filling a gap: giving organizations outside China an open-weight option they can self-host without depending on a Chinese lab’s model or a US closed-model subscription.

Sources: Mistral’s own announcement (“Introducing Mistral Large 4,” mistral.ai/news/mistral-large-4) and model/pricing documentation (docs.mistral.ai); Reflection AI’s own announcement (“Introducing Beam,” reflection.ai/blog/introducing-beam); Mistral’s funding announcement (mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier) and reporting from TechCrunch, France 24, and Value Add VC on the Series D; TechCrunch’s reporting on Beam and Reflection AI’s funding history; third-party benchmark cross-checks via cellcog.ai, read October 6–7, 2026.

Researched and drafted with AI assistance from the sources linked above, and published by MeasuredRun. How we use AI.