Reflection's Beam bets on fewer tokens, not higher scores
Beam is a 501B-parameter model with Apache 2.0 weights promised this month. Its pitch is not that it wins benchmarks, but that it spends less compute getting close.
- Model
- Beam
- Maker
- Reflection AI
- Size
- 501B total, 23B active
- Context
- 1M (256K during RL training)
- License
- Apache 2.0 (promised)
- Weights
- Promised
Most model launches lead with the benchmark they win. Reflection AI’s announcement of Beam does something less common: it publishes a table in which Beam mostly loses, and argues you should care anyway.
Beam was announced on October 5 as an open-weight model. The weights, a technical report, and a model card are promised under Apache 2.0 later this month. Until then, access is through a waitlist.
How it works
Beam is a sparse mixture-of-experts model: 501 billion parameters in total, of which 23 billion are active for any given token. It is text-only.
Reflection says it pretrained on 23.8 trillion tokens, then ran four weeks of reinforcement learning on about 10,500 NVIDIA GB300 GPUs. That phase produced more than 100 million attempts at tasks across roughly a million environments. The RL runs used up to 256K tokens of context, and a separate training stage extends the effective context to 1M.
Safety training was handled in an unusual way. Reflection trained a second model from the same starting checkpoint with its own safety and alignment pipeline, then distilled both “teachers” into the final model. The safety evaluations themselves are due with the technical report.
The efficiency claim
The headline is that Beam scores about the same as GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute.
It is worth reading how that is measured. Reflection estimates compute as twice the active parameter count times the average number of tokens the model generates per attempt. That leaves out reading the prompt, attention over long contexts, and everything about actually serving the model. Reflection itself calls it an approximate comparison, not a measured cost.
So the claim is narrower than it sounds: Beam has fewer active parameters than its rivals and writes fewer tokens to reach an answer. Whether that shows up on your bill depends on hardware and workload.
The benchmarks
All Beam scores are Reflection’s own. The post was revised on October 8, and these are the updated figures.
| Benchmark | Beam | GLM 5.2 | GLM 5.3 | Kimi K3 | Qwen 3.8 Max | DeepSeek V4.1 Flash |
|---|---|---|---|---|---|---|
| Terminal Bench v2.1 | 80.1 | 81.0 | 88.2 | 88.3 | 86.6 | 90.6 |
| DeepSWE v1.1 | 44.4 | 44.0 | 61.0 | 68.0 | 51.0 | 74.2 |
| SWE Bench Pro v1 | 65.5 | 62.1 | n/a | n/a | 67.7 | n/a |
| GPQA Diamond | 90.5 | 91.2 | 91.7 | 93.5 | 92.6 | 90.9 |
| Humanity’s Last Exam, no tools | 36.2 | 40.5 | 42.3 | 46.9 | 43.6 | 39.1 |
| AA-LCR (long context) | 79.3 | 78.3 | 79.7 | 88.7 | 80.3 | 84.0 |
The pattern is consistent. Beam trades blows with GLM 5.2, the previous GLM release, and sits behind GLM 5.3, Kimi K3, and DeepSeek V4.1 Flash on most rows. In the table, n/a means Reflection did not report a score. Reflection says as much, describing Kimi K3 as still ahead on raw capability.
One detail from the write-up, noted by Help Net Security: Beam’s browsing scores rose during a training phase that contained no browsing tasks, and with web access it went looking for other language models to query. That is worth knowing before you give it a network connection.
Try it yourself
You cannot yet. Early access is by waitlist at Reflection’s platform site, and the weights are not public. We have not run Beam ourselves: this is based on Reflection’s announcement and Help Net Security’s reporting. It is listed as promised on the Open Weights tracker until there is something to download.