Overseas open-source models pick up speed: Mistral and 'American DeepSeek' unveil new models

On October 5, 2026 local time, the American startup Reflection, dubbed the "American DeepSeek," released its first open-weight model Beam; one day later, France's Mistral launched its largest model to date, Mistral Large 4. The two companies' primary comparison targets at launch were highly consistent, all being GLM, Kimi, DeepSeek, and Qwen. This choice of benchmarks itself shows that Chinese open models have become the default coordinate by which Western peers measure their own position.

Reflection was founded in 2024, and both founders came from Google DeepMind. According to Reflection's founders, what truly changed the company's direction was DeepSeek V3: after Chinese open models advanced rapidly while the West long failed to produce an equivalent alternative, the company decided to train frontier open models itself and gradually made its goal explicit as "bringing the open-model frontier back to the West." Two years later, Beam represents Reflection's first real delivery.

Mistral faces greater pressure. Two or three years ago it was once the most representative company in open models, but as DeepSeek, Qwen, Kimi, GLM and others iterated rapidly, Mistral's presence on frontier capability leaderboards steadily declined. This September, CEO Arthur Mensch was directly asked whether Mistral's models had fallen outside the top 20 in Artificial Analysis rankings; he responded that a new generation of large models was imminent, and that the company had already planned its next two generations of products for the coming year. Two weeks later, Mistral Large 4 appeared.

Beam is an MoE model with 501 billion total parameters and 23 billion activated parameters per token, primarily targeting coding, reasoning, and agent workloads. The model was pretrained on 23.8 trillion tokens, and Reflection then used 10,500 Nvidia GB300s for 4 consecutive weeks of high-compute reinforcement learning, generating over 100 million rollouts. The model is still in its final red-teaming stage, and Reflection plans to publicly release the model weights, technical report, and development tools in October.

From the test results Reflection itself published, Beam's biggest highlight is not absolute capability but inference efficiency. The company claims that on some advanced reasoning tasks, it can achieve near-comparable performance using roughly one-quarter to one-third of GLM-5.2's inference compute.

But when compared horizontally with China's latest-generation models, Beam still has a clear gap before it can truly "catch up." On the agent coding test DeepSWE v1.1, Beam scored 44.4, nearly on par with GLM-5.2's 44.0, but below GLM-5.3's 61.0, Kimi K3's 68.0, and DeepSeek V4.1 Flash's 74.2; on Terminal Bench v2.1, Beam scored 80.1, while GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash reached 88.2, 88.3, and 90.6 respectively; on Humanity's Last Exam (HLE), Beam scored 36.2, also below GLM-5.3's 42.3 and Kimi K3's 46.9.

Therefore, Beam currently pulls American open models back near China's top models, but what it has mainly caught up with is the previous generation GLM-5.2, still about a generation behind China's latest round of open models. Reflection itself does not shy away from this: it officially acknowledges that Kimi K3 still leads in absolute capability, and shifts Beam's selling point to "how much intelligence you can get per unit of inference compute." CEO Misha Laskin positions Beam as a "workhorse model" that enterprises can run long-term, rather than simply the largest model chasing leaderboard rankings.

This also resembles part of the logic that initially won DeepSeek over the market: the value is not in whether the model is number one, but in how much money it costs to reach a certain capability. The difference is that Reflection can hardly be called a "low-cost version of DeepSeek." Beam's final round of reinforcement learning alone simultaneously occupied more than 10,000 of the latest GB300s; the company has already raised billions of dollars, received a large investment from Nvidia, and locked in substantial compute through partnerships with SpaceX, Nebius, and others. It is a "heavy-asset version of DeepSeek" backed by American capital and Nvidia's latest hardware.