Reflection AI introduced Beam on October 5, and on paper it is another huge mixture-of-experts model with billions of parameters: 501B total, 23B active.
What makes Beam interesting is not that Reflection is claiming it will top every benchmark. The company is putting much more emphasis on how efficiently the model uses tokens for the level of performance it delivers.
Artificial Analysis, which received early access to Beam, says the first signs are promising and that it could end up being one of the most token-efficient open models in its intelligence range. For now though, that is still only an early impression. Beam has not appeared on the Artificial Analysis leaderboard yet, so there is no public result to back that claim up. One caveat before the numbers: every Beam figure below comes from Reflection’s own launch material. The model is still in final red-teaming, and the weights are due later this month under Apache 2.0.
Another huge MoE
Every few days, a new open-weight model arrives with a bigger number attached. Mistral Large 4 landed a day later: a 1T model with 49B active. Beam has about half the total size, a much smaller active share, and weights that arrive later.
| Model | Total | Active | Weights | DeepSWE v1.1 |
|---|---|---|---|---|
| Mistral Large 4 | 1T | 49B | End of October | 62% |
| Reflection Beam | 501B | 23B | This month | 44% |
In Mistral’s own comparison, Beam sat at the bottom of the chart, below GLM-5.3, DeepSeek V4 Pro and Qwen 3.8 Max. On raw score, Beam is not the model to beat.
The difference
Beam’s main argument is intelligence per token. In Reflection’s own HLE chart, as reproduced by DataCamp, Beam scores higher than two rivals while using far less compute.
| Model | HLE score | Compute per attempt |
|---|---|---|
| Reflection Beam | 35.3% | about 0.82 PFLOP |
| Inkling | about 26% | about 4.1 PFLOP |
| Nemotron 3 Ultra | about 21% | about 1.9 PFLOP |
That works out to roughly 5x less compute than Inkling and about 2.3x less than Nemotron 3 Ultra, while still posting a higher score. Reflection also claims Beam can match GLM-5.2 on advanced reasoning while using around 3 to 4x less inference compute. GLM-5.2 has roughly 40B active parameters, compared with Beam’s 23B.
There are a few catches, though. The chart compares Beam against two of the weaker models in Reflection’s table. On HLE without tools, Beam scores around 36, while GLM-5.2 reaches 40.5, Qwen 3.8-Max hits 43.6, and Kimi K3 reaches 46.9. So when Reflection says Beam “matches GLM-5.2,” that is still its own framing, and the exact numbers are not perfectly consistent between the company’s charts and tables.
The compute figure is also only an estimate. Reflection calculates it using roughly twice the active parameter count multiplied by the average number of tokens generated per attempt. That is useful for comparison, but it does not include prompt prefill, attention cost, or serving overhead. Those can become a much bigger part of the workload in long-context agent tasks, and the formula naturally makes models with fewer active parameters look better.
There is also a difference between token efficiency and compute efficiency. Artificial Analysis is talking about how many tokens Beam uses, while Reflection’s chart is estimating how much compute those generations require. They are related, but they are not measuring the same thing.
Both still matter a lot for coding and agent workloads, where a single task can involve thousands of model calls. That is also where Beam has one of its strongest results, scoring 80.9 on SWE-bench Verified compared with 77.6 for Inkling and 70.7 for Nemotron 3 Ultra.
What you give up
Beam sits around the same level as GLM-5.2 but still falls behind the strongest open-weight models. These results come from Reflection’s own testing.
| Benchmark | Beam | GLM 5.2 | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 80.1 | 81.0 | 88.3 | 90.6 |
| DeepSWE v1.1 | 44.4 | 44.0 | 68.0 | 74.2 |
Mistral Large 4 scores 17 points higher on DeepSWE, but it also activates more than twice as many parameters, with 49B active compared with Beam’s 23B. If raw performance matters most, the larger model has the advantage. Beam is aimed more at workloads where you are running lots of tasks and want to keep the cost of each one lower.
What you get
Beam is text-only and released under Apache 2.0. Reflection claims a 1M-token context, while the beta API documents 256K.
What 501B means at home
It means you are not running it, unless you own a home server. Active parameters set compute per token, but total parameters set memory, and all 501B still have to be loaded.
| Quant | Weights alone |
|---|---|
| Q8_0 | about 530 GB |
| Q4_K_M | about 300 GB |
| 2-bit | about 125-165 GB |
These are estimates from arithmetic, before the context cache. Even the smallest row is a long way from an 8 GB card.
Conclusion
Beam is not aiming to beat every open model on raw performance. Its goal is getting close while using a lot less compute. For now, that argument mostly comes from Reflection’s own estimates and an early indication from Artificial Analysis that has not been published as a full result yet. The real test starts once the weights are out and independent benchmarks can show whether those efficiency claims actually turn into lower costs and faster runtimes.
