Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam8G
- Frontier reasoning efficiency
- Advances the Western open frontier on coding & agentic tasks
- Trained end-to-end from scratch
Full weights release this month.
Learn more about Beam8G
565
1,100
8,601
2,514,057
We scaled Beam with the largest publicly documented RL run we’re aware of, which ran on 10.5k GB300s for 4 weeks.
Algorithmic advances coupled with distributed infra enabled an RL system that scales.
Across our eval suite, capabilities continued to improve as we increased RL - u.
Algorithmic advances coupled with distributed infra enabled an RL system that scales.
Across our eval suite, capabilities continued to improve as we increased RL - u.
Oct 5, 2026 · 7:11 PM UTC
11
26
446
58,261
Sustained RL improvement was made possible by a strong foundation for reasoning.
Beam was pretrained in 4 weeks on 24T high-quality tokens to have innate coding capabilities.
Innovations in MoE stability and large-scale data curation & deduplication enabled us to produce a founavailable open-source base models of the same class.
Beam was pretrained in 4 weeks on 24T high-quality tokens to have innate coding capabilities.
Innovations in MoE stability and large-scale data curation & deduplication enabled us to produce a founavailable open-source base models of the same class.
5
14
332
37,321
Beam combines strong agentic performance, efficient reasoning, and a 500B form factor to give enterprises, governments, and developers a true workhorse open model. It's our first step toward building frontier intelligence that is open and widely accessible to all.
Now in the final stages of red-teaming, Beam will be released under an Apache 2.0 license this month, along with quantized FP8 and NVFP4 numerics for efficient deployment.
Sign up for early access: platform.reflection.ai
Now in the final stages of red-teaming, Beam will be released under an Apache 2.0 license this month, along with quantized FP8 and NVFP4 numerics for efficient deployment.
Sign up for early access: platform.reflection.ai
15
21
439
31,072
@reflection_ai @grok estimate costs of this based on the rate of GB300 to rent in September 2026. Taken into account economies of scale and other real world factors.
1
0
2
388
@reflection_ai I hate to ask but, yall got any GB300s I can get? My lord that's a truckload!
0
0
0
62
@reflection_ai 80M rollouts? hopefully not an error? mimo had 753k rollouts and did 72 on deepswe 1.1, if my memory serves right.
0
0
2
219
@reflection_ai The scale of that RL run is absolutely massive. We actually went deeper on the technical details of Beam here: x.com/AGTPinsights/s…
Reflection just unveiled Beam, its first open model, today. Here's what you need to know.
Yesterday's Axios report is now confirmed. Reflection announced Beam on Oct 6, 2026. It is a text-only mixture-of-experts model with 501B total parameters and 23B active. It is released cense. Full weights and technical details are due this month, with distribution through hyperscalers and neoclouds.
Beam is built for coding, agentic and scientific work. It was trained end-to-end from scratch. Pretraining took 4 weeks on 24T tokens. Reflection says it ran the largest publicly documented RL run it knows of, on 10.5k GB300s for 4 weeks. Capabilities kept improving as RL scaled.
Reflection says Beam is 3-4x more efficient at inference than GLM 5.2, and over 4x more efficient than leading Western open models. That means tasks finish faster and cheaper. GLM 5.2 is Z.ai's open model, with roughly 744B total parameters.
On Reflection's own benchmarks, Beam is competitive with GLM 5.2. Investor Stephanie Zhan says it approaches Qwen 3.8 Max at about 1/4 the size. Reflection also reports Beam beats Inkling on four coding tests where both report results. Inkling is multimodal.
Key numbers:
- 501B total parameters, 23B active
- 24T pretraining tokens, 4 weeks
- RL run on 10.5k GB300s, 4 weeks
- 44 on DeepSWE
- 80 on Terminal-Bench 2.1
- 15 on CritPt
- 3-4x more efficient than GLM 5.2
Reflection pitches Beam as the American answer to China's open model lead. Ollama says more US models are coming to its platform.
0
0
0
190
@reflection_ai does per-token checkpoint tracking feed into importance weights, a policy-lag cutoff, or both?
0
0
0
49