America's big open-model hope is here, and its own chart has it behind four Chinese models
Reflection AI unveiled Beam on October 5, a 501-billion-parameter open model. The company's own benchmark table shows where it trails. Its pitch is cost per answer, and the weights aren't out yet.
Reflection AI put out its first model on October 5, and the company’s own blog post is more candid than most launch posts. Beam, it says, is “competitive with larger open models like GLM 5.2” and “approaching Qwen 3.8-Max” on coding. Then comes the sentence I’d have cut if I were its marketing team: frontier open models like Kimi K3 “remain ahead on raw capability.”
So this isn’t the launch of a Chinese-model killer. It’s a bet on something else.
What Beam is
Beam is a mixture-of-experts model with 501 billion parameters in total, of which about 23 billion switch on for any one token. Reflection, a New York company founded in 2024 by former Google DeepMind researchers, says it trained on 23.8 trillion tokens, then spent four weeks of reinforcement learning on 10,500 Nvidia GB300 chips. The company says that run produced more than 100 million rollouts.
Fortune reports that Reflection has a $25 billion valuation, a $6.3 billion compute deal with SpaceX, and supply agreements with the US Energy and War departments.
One catch matters more than the specs. You can’t download Beam today. Reflection says it’s in final red-teaming and promises the weights under an Apache 2.0 licence “later this month”, plus a technical report and model card. Until those land, every number below is the company’s own.
Reading the company’s own table
Reflection’s post compares Beam with seven other open models. Two tests have scores for nearly everyone:
| Test | Beam | GLM 5.2 | GLM 5.3 | Kimi K3 | Qwen 3.8 Max | DeepSeek V4.1 Flash |
|---|---|---|---|---|---|---|
| DeepSWE v1.1 | 44.4 | 44.0 | 61.0 | 68.0 | 51.0 | 74.2 |
| Terminal Bench v2.1 | 80.1 | 81.0 | 88.2 | 88.3 | 86.6 | 90.6 |
Beam sits with GLM 5.2, which Z.ai released in July. Since then Z.ai has shipped GLM 5.3, and that newer model is well ahead on both rows. Several of Beam’s strongest rows, like SWE-Bench Verified at 80.9, have blanks for the Chinese models, so there’s nothing to compare against. Those are vendor-run numbers, not a third party’s.
I’m not saying the chart is dishonest. Reflection labels the gaps “NR”, not reported. But a reader skimming the headline would never guess that DeepSeek’s Flash model scores 74.2 where Beam gets 44.4.
The real pitch is cost
Reflection’s case is that Beam reaches GLM 5.2’s level of reasoning while using three to four times less inference compute. It’s measuring that by multiplying active parameters by the tokens generated per attempt, and it says outright this is an estimate that leaves out serving overhead. It isn’t a price list.
If it holds up, though, it’s a real selling point. Companies that want a model they can run on their own servers care about the bill per task as much as the leaderboard. And a US-made option matters to buyers who can’t or won’t use Chinese models, a group that includes the government customers Reflection already has.
Mistral showed up the next day
Europe answered inside 24 hours. On October 6, Reuters reported, Mistral unveiled Mistral Large 4, nicknamed Le Chonk, and CEO Artur Mensch said it beats Chinese models “on certain aspects, including cyber” without naming which models or benchmarks. It goes public on October 27. Mensch’s claim has no numbers behind it yet, so it deserves the same caution as Beam’s.
Where I land
I’d wait for the weights. An open model is the one kind of release anybody can check, and people will run Beam on their own tests within days of the download going live. If its efficiency claim survives that, Beam becomes the default Western pick for companies that want open weights. If not, it’s a well-funded model that lands one generation behind.
The safety side is worth watching too. We just covered an open Chinese model that built working exploits almost as well as a rationed Anthropic one, and Reflection says it will publish its own safety evaluations and open-source the tests. I’d read those before the benchmark table.
Sources
- Introducing Beam: Reflection’s 501B open-weight model (Reflection AI, October 5, 2026)
- Nvidia-backed Reflection AI unveils its first open model, Beam (Fortune, October 5, 2026)
- France’s Mistral launches AI model it says outperforms some Chinese rivals (Reuters, via MarketScreener, October 6, 2026)