Two open-weight giants land: Mistral Large 4 and Reflection Beam

Mistral Large 4 is in public preview. It's a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 49B active. It was introduced on Oct 6 and is available through the API on Mistral Studio. Mistral says the weights follow by the end of October, after more red-teaming work.

Reflection AI's Beam came a day earlier. It's a 501B-parameter sparse MoE with roughly 23B active per token, built for coding, reasoning and agentic tasks. Reflection claims it matches leading Chinese open models on reasoning at 3-4x less inference compute. That's a vendor claim, not an independent benchmark. A beta API is live, with Apache 2.0 weights promised later this month.

Both launches are previews. Neither can be self-hosted, quantized or fine-tuned until the weights ship.

Why it matters: Western labs are finally contesting the open-weight frontier that Chinese models have dominated, but only if the weights actually drop on schedule.