The US finally has an answer to DeepSeek and Qwen
For months, the open-weight AI race has belonged to China. Models like DeepSeek, Qwen, and Z.ai's GLM series have been faster, cheaper, and surprisingly capable — while Western labs mostly kept their best systems behind paid APIs. On October 5, 2026, a small Brooklyn startup called Reflection AI stepped in to change that story with the launch of Beam, its first open-weight model.
The headline number is 501 billion. Beam is a mixture-of-experts model with 501 billion total parameters, of which only 23 billion activate for any single task. That sparse design is the whole point: instead of running the entire model every time, Beam lights up just the part it needs, making it much cheaper and faster to operate.
The benchmark claims — and the big asterisk
Reflection's own technical report says Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute. For perspective, GLM-5.2 is a much bigger machine: roughly 744 billion total parameters with 40 billion active. Reflection also says Beam is closing in on Alibaba's Qwen3.8-Max on coding and agentic tasks, and it scores 80.9 on SWE-Bench Verified and 78.0 on SWE-Bench Multilingual — both serious coding tests.
The honest asterisk: these numbers come from Reflection itself. Independent researchers have not verified them yet. Open-model launches often look great on company-chosen benchmarks, so treat the claims as promising rather than proven.
How you build a model like this
The training details are unusually specific — and they read like a flex. Beam was pretrained on 23.8 trillion tokens using 6,144 Nvidia GB300 chips, finishing the run in under four weeks with what the company calls 92.3% "goodput," meaning almost all of that GPU time went into useful work rather than failures and restarts. Then came the hard part: over 100 million reinforcement-learning rollouts across roughly a million synthetic coding, agentic, and science environments, using another 10,500 GB300 chips.
That scale explains Nvidia's enthusiasm. The chipmaker put $800 million into Reflection's last funding round, and it is clearly betting that startups like Reflection prove out its "AI factory" pitch: sell the chips, then fund the labs that show what the chips can do. Reflection has reportedly committed billions in compute through 2029, and it even signed a deal with SpaceX for extra capacity at the Colossus 2 data center.
Why "open" is the real story here
Beam is not just another model — it is an open-weight one. That means companies will be able to download the weights, run the model on their own hardware, and fine-tune it on their own private data, instead of paying API bills and sending data to someone else's servers. Full weights and the technical report are due later this month, with early access already available through the company's developer platform.
This matters for a practical reason. Right now, the most capable open models come from Chinese companies, and many Western firms — plus parts of the US government — would prefer a Western alternative they can inspect and control. Beam is positioned exactly there: cheaper than rivals, aimed at coding and AI agents (the workloads enterprises actually want), and open enough to build on.
The competition is not standing still. Reflection's most direct American rival may be Inkling, the open model from Mira Murati's Thinking Machines Lab, while Moonshot's Kimi K3 reportedly still leads on raw capability. Open-source kingmakers like Meta's Llama team, Mistral, and Cohere are all in the same fight.
If Beam's efficiency claims hold up in the real world, it could do something rare: give enterprises Western open weights without the painful cloud bills. The weights release later this month is the moment of truth — that is when independent developers can finally test whether the hype matches reality.
Do you think open models will win the AI race, or will the closed giants stay ahead? Let us know in the comments.
Comments
Post a Comment