The West’s open-weight flagships are announced, not shipped

On Monday, Reflection announced Beam. On Tuesday, Mistral announced Large 4, nicknamed "Le Chonk." Each calls its model the strongest open-weight model built outside China. Neither has published its weights.

What was announced

Beam is a 501-billion-parameter mixture-of-experts model that activates 23 billion parameters per token. It is text-only and built for coding and agentic work. Reflection says the weights come "later this month" under Apache 2.0, and early access is a waitlist while red-teaming finishes. It is Reflection's first open-weight model.

Le Chonk has 1 trillion parameters and 49 billion active. It takes images and text in and returns text. A public preview runs on Mistral's API today, trained on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data centers. Weights follow at the end of the month. Mistral told press October 27, while the Hugging Face placeholder page counts down to October 31.

Same crown, thinner proof

Reflection does not claim to lead. Its own post says Kimi K3 "remain[s] ahead on raw capability," and pitches Beam on efficiency: reasoning scores comparable to Zhipu's GLM 5.2 with 3–4× less inference compute, by Reflection's own estimate. In its table, Beam scores 80.1 on Terminal Bench v2.1. GLM 5.2 gets 81.0, Kimi K3 88.3, and DeepSeek V4.1 Flash 90.6.

Mistral makes the bigger claim: strongest open-weight model outside China "by a substantial margin," and state of the art among open models on cybersecurity, finance, and law. Both sets of numbers come from the companies. Independent testing has to wait for downloads.

The gap is narrowing, not closed

Bloomberg Intelligence reported this week that top Chinese models now trail US rivals by about 3% on benchmark scores, down from about 9% in May and 15% earlier this year, after DeepSeek's V4.1 Flash. That compares all models, and the leading US ones are closed. It is not an open-weight gap. On Reflection's own table, Chinese open models still lead the open tier.

Announced is not released

We made the same point in our August open-weight report: Glimmer was real and downloadable, Spark was a promise. Until the files land, nobody can self-host, read the license, or run their own evals. At 16 bits, 501 billion parameters is roughly a terabyte of weights, so the hardware question comes before the download.

What to watch

Whether Beam ships under a clean Apache 2.0 file, and whether Mistral meets October 27. Then independent results against Kimi K3, GLM 5.3, and DeepSeek V4.1 Flash. The broader stack, from chips to distribution, is in our Chinese AI report.

The point

Two Western labs now say they have the best open model outside China. Both are still waiting on the one thing that makes a model open: the weights.