Reflection's Beam Has 501B Parameters and 23B Active, and Its Weights Are Still Due "Later This Month"
A.I. / news
Reflection's Beam Has 501B Parameters and 23B Active, and Its Weights Are Still Due "Later This Month"
The company's own benchmark table shows Kimi K3, GLM and Qwen ahead on the rows below, and its pitch rests on a vendor-supplied claim of three to four times less inference compute.

Reflection AI announced Beam on October 5, 2026, a 501-billion-parameter open-weight language model that activates 23 billion parameters per token. The weights are not downloadable yet. The company said it will release them under an Apache 2.0 licence "later this month".
The announcement is a post on Reflection's own blog. It names no individual researcher and gives no price for the early-access API, which sits behind a waitlist at platform.reflection.ai. The post says the model is still undergoing final red-teaming and evaluations.
What Beam is, by the numbers
Beam is a sparse mixture-of-experts model. Reflection said it was pretrained on 23.8 trillion tokens drawn from the web and from proprietary licensed datasets, on 6,144 Nvidia GB300 NVL72 GPUs in under four weeks, at 92.3% goodput.
Reinforcement learning ran for four weeks on 10.5K GB300 GPUs. The post counts more than 100 million rollouts across roughly 1.3 billion sandboxes, drawn from one million coding, agentic and STEM tasks.
The context window was 256K tokens during reinforcement learning and was extended to 1M tokens in midtraining. Reflection did not say what hardware is needed to serve the model at either length.

Beam against the models in Reflection's own table
Reflection's headline claim is efficiency. The post says Beam reaches scores comparable to Z.ai's GLM-5.2 while using three to four times less inference compute. That is a vendor-supplied claim, and the post does not name an outside party that reproduced it.
The benchmark tables compare Beam with Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8-Max and DeepSeek V4.1 Flash. On the rows below, a rival scores higher than Beam in every case. The post says as much, adding that frontier models such as Kimi K3 keep an advantage in raw capability.
| Benchmark | Beam | Highest rival in the table |
|---|---|---|
| DeepSWE v1.1 | 44.4 | 68.0, Kimi K3 |
| SWE Bench Pro v2-Hard | 77.2 | 84.3, GLM 5.3 |
| Terminal Bench v2.1 | 80.1 | 90.6, DeepSeek V4.1 Flash |
| AIME 2026 | 97.8 | 99.2, GLM 5.2 |
| GPQA Diamond | 90.5 | 93.5, Kimi K3 |
| MCP Atlas | 78.7 | 84.5, Qwen 3.8-Max |
| AA-LCR | 79.3 | 88.7, Kimi K3 |
| LongBench v2 | 65.5 | 66.3, Qwen 3.8-Max |
All figures are Reflection's own, taken from its post on October 5. The rival column shows the best score among the other models in that table for each row.
Who is behind it, and what is not yet public
Reflection's chief executive is Misha Laskin and its chief technology officer is Ioannis Antonoglou, according to Startup Fortune and Sources. Sources reports a $25 billion valuation and $4.6 billion raised, with Nvidia, Sequoia and Lightspeed among the investors. Startup Fortune puts the raise at about $4.7 billion. The two outlets differ by $100 million and the company post does not settle it.
Startup Fortune also notes that the benchmark figures have not been independently verified. Reflection's announced licence is permissive, but the post gives no licence for the training data, and "proprietary licensed datasets" is the only description of that portion.
Other open-weight releases have shown how licence terms can diverge from the headline. The Qwen3.8-27B release kept the larger flagship under a different licence, and H Company's Holo4-27B carries a non-commercial bar on an Apache 2.0 base. The Beam model card, due with the weights, is where the licence file will have to be checked.
Reflection said Beam is the first model in a series and that it is already training the next one. The weights, technical report and model card are due later in October, with no day given.
Sources
More in A.I.
- 01Thinking Machines Ships Inkling 2 Mini, a 26.4B Apache 2.0 Model Sized for One DGX SparkThe model card lists 3.4B active parameters, a 1M-token window and a 20GB checkpoint, but does not say who ran most of its benchmark rows.
- 02Anthropic Closed Its Debut at $97.17, 17% Over the $83 IPO Price, as Nasdaq Fell 2%The $81.34 billion offering opened 34% higher, then gave back most of the gain while Nvidia, AMD and Micron dropped.
- 03Google's Gemini Tool Retrieval Preview Adds 26 Tokens Per Idle Tool and Publishes No Savings FigureAnthropic and OpenAI each reported token reductions for the same defer_loading flag. Google's November 19 post reports none.
- 04OpenAI Says New Safeguards Cut Reasoning Extraction From 54.0% to 0.2%The figures are from OpenAI's own red-team evaluation, and the post names DeepSeek and Moonshot AI as distillers of U.S. models.