Mistral and Reflection AI Launch Open-Weight Models to Challenge China, Claims Not Yet Independently Verified

Two Western AI companies made their move against China's open-weight AI lead this week.

Reflection AI, a US startup backed by Nvidia, introduced its first open-weight model, Beam, on Monday, October 5. It's a 501-billion-parameter mixture-of-experts system with 23 billion parameters active at any given time. The weights aren't out yet. They're due later this month, according to Euronews. Until then, Reflection's performance claims sit unverified.

Mistral AI, the French company, followed a day later. On Tuesday, October 6, it launched a public preview of Mistral Large 4, internally called ML4 and nicknamed "Le Chonk." It's a 1-trillion-parameter model with 49 billion active parameters, trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months inside Mistral's own European data centers, according to Quartz. The full weights go public on October 27.

What Mistral is claiming, and what it isn't

Mistral says Le Chonk will rank among the top open-weight models globally and will be, by a substantial margin, the strongest open-weight model built outside China, per Quartz. The company is more careful about China specifically. Its own launch blog post, as reported by Trending Topics, says Large 4 is competitive with the strongest open models worldwide but does not claim a clean win over China's best.

The margins against Chinese rivals are thin where Mistral does claim an edge. On the DeepSWE benchmark, Large 4 edges out Z.ai's GLM-5.3. On FinWorkBench, it ties DeepSeek's V4 Pro, according to Trending Topics. In a blind coding evaluation, Large 4 came in second of five models, behind Anthropic's closed Claude Opus 5. Mistral itself acknowledges the model still trails in coding relative to other frontier systems, Quartz reported.

None of this has been independently confirmed. Trending Topics noted there are no independent benchmarks based on the open weights yet, because the weights aren't out until October 27. The same caveat applies to Reflection's Beam, which Euronews and Yahoo both noted has performance claims that "have not been independently verified."

Why cybersecurity is the pitch

Mistral is leaning hardest on cyberdefense. Guillaume Lample, the company's co-founder and chief scientist, said the model will let "enterprises and governments... defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks." Mistral says the model ranks among the top five globally on the Artificial Analysis Cyber Index, an outside evaluation, per Trending Topics.

That's a real concern worth taking seriously on its own terms. Open-weight models can be downloaded, modified, and run by anyone, including bad actors, which is exactly why Axios flagged that "the same characteristic that gives customers control over an open-weight model also limits the developer's control over how it is used." A model built for defense can be repurposed for attack once it's off a company's servers.

Mistral's counter is that openness is itself a safety feature, letting researchers and governments see inside the system rather than trusting a black box. Pierre Stock, Mistral's first employee and VP of science, told Euronews that American AI labs are incentivized to "market rogue events" involving AI agents acting on their own, because it supports a narrative that "only a few trusted players can take care of the technology." He called it a "monopoly... controlled by the Americans," arguing open-weight models break that up. OpenAI and Anthropic were not quoted responding to this assertion in these reports.

The money behind the comeback

Mistral's timing isn't an accident. The company raised €3 billion ($3.4 billion) in a Series D round led by Samsung Electronics last month, at a valuation north of €21 billion (roughly $24 billion), which Mistral calls the largest equity round ever completed by a European tech company, per Quartz. That followed an $830 million debt raise for a data center outside Paris, with a second facility planned in Sweden as Mistral targets 200 megawatts of combined European capacity by the end of 2027.

Wired reported Mistral's earnings have reportedly grown twentyfold over the past year, though that figure comes from the company rather than an audited filing.

Large 4 was originally slated for a summer release and slipped, according to Trending Topics, with its predecessor already nearly a year old. The preview is priced at $1.36 per million input tokens and $4.18 per million output tokens through Mistral's API, aimed at banks, logistics firms, defense contractors, and government agencies running the model in their own data centers.

What's still unresolved

When Mistral's weights go public on October 27, outside researchers can finally run their own benchmarks instead of relying on the company's internal comparisons. Reflection's Beam faces the same reckoning once its weights ship later this month. Until then, both companies' claims against China's GLM-5.3, DeepSeek V4 Pro, and Alibaba's models remain exactly that: claims.