Reflection's 501B Beam Model Gets Apache 2.0 Weights on Hugging Face
A.I. / news
Reflection's 501B Beam Model Gets Apache 2.0 Weights on Hugging Face
Hugging Face's launch post lists three checkpoints, a 1 million-token context window and vendor-run benchmarks that differ from the figures Reflection published at announcement.
Reflection released the weights for Beam, a 501-billion-parameter mixture-of-experts model, under the Apache 2.0 license on Friday, Oct. 16, according to Hugging Face's launch post. A mixture-of-experts design routes each token to a subset of the model's parameters instead of all of them.
Reflection announced Beam earlier in October. LetsDataScience dates the announcement Oct. 5 and VKTR dates it Oct. 7. Reflection said then that the weights, a technical report and a model card would follow "later this month," and that Beam was in final red-teaming with access through a waitlist.
What the three Beam checkpoints need
Beam has 23 billion active parameters per token. Hugging Face describes 52 layers, 256 experts per MoE layer, eight experts routed per token and no shared expert. Of the 52 layers, 43 use a sliding attention window of 512 tokens and nine attend globally.
Hugging Face puts the key-value cache at full context at about 36.08 GiB per sequence, against 208 GiB if every layer were global.
| Checkpoint | Size | Hardware named in the post |
|---|---|---|
| BF16 | about 1 TB | 16 H200s with transformers or vLLM |
| FP8 | not given | 8 H200s with SGLang |
| NVFP4 | about 280 GB | 4 B300s with SGLang |
The post calls NVFP4 the most practical option and says it runs natively on Blackwell GPUs. Its planned setup is four B200s with 720 GB combined. Other listed options are an 8xB200 node, two B300s or 8xRTX Pro 6000. For a comparison with a model sized to fit one 24 GB card, see the Terminal's coverage of Nvidia's Nemotron 3.5 Bolt.
The Beam repository on Hugging Face returned an authorization error (HTTP 401) when this article fetched it. Whether download requires a login or accepting a form is not confirmed here.
Every benchmark is Reflection's own
The scores in Hugging Face's post were run by the Reflection team, per the post. On the BF16 checkpoint they are 81.6 percent on SWEBench Verified, 97.7 percent on AIME 2026, 90.6 percent on GPQA Diamond, 80.1 on Terminal Bench v2.1 and 88.3 percent on LiveCodeBench. SWEBench Verified falls to 80.2 percent on FP8 and 80.7 percent on NVFP4.
The figures from the announcement do not line up with those. LetsDataScience reported 80.9 on SWE Bench Verified and 77.2 on SWE Bench Pro v2-Hard from Reflection's own post. VKTR reported 65.5 on SWE Bench Pro v1. The Terminal Bench v2.1 score of 80.1 matches in both. This article cannot say whether the 80.9 and 81.6 reflect different harness settings.
VKTR's table, which it attributes to Reflection, places Beam behind three rivals on Terminal Bench v2.1.
- Kimi K388.3 points
- Qwen 3.8 Max86.6 points
- GLM 5.281 points
- Beam80.1 points
Source: VKTR, reporting Reflection's own results, accessed 2026-10-16
Reflection said Beam matches GLM 5.2 on advanced reasoning while using three to four times less inference compute, and acknowledged that Kimi K3 remains ahead on raw capability. That efficiency claim is vendor-supplied. Artificial Analysis, which received early access, wrote that "early indicators suggest Beam will be one of the most token-efficient open models we've seen for its level of intelligence," per VKTR.
Training claims and the Nvidia talks
Reflection said pretraining covered 23.8 trillion tokens on 6,144 Nvidia GB300 GPUs in under four weeks, followed by a four-week reinforcement learning run on 10,500 GB300s that generated more than 100 million rollouts. LetsDataScience describes the pretraining data as web and licensed datasets. The model card text could not be retrieved, so what the licensed data is has not been checked.
The Financial Times reported on Oct. 10 that Nvidia is in early talks to deepen its investment in Reflection or acquire it, according to LetsDataScience's summary. Reuters said it could not verify the report, and neither company responded to Reuters. Nvidia led Reflection's $2 billion round in October 2025, investing $800 million at an $8 billion valuation.
For scale on the same hardware family, Microsoft's MAI-Base-2 trained on 16,000 GB200s in 38 days.
Reflection has not yet published the technical report it promised for later in October. That is the document that would say what the 23.8 trillion tokens contain.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark TablesOpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
- 03Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 04StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.