Reflection Announces Beam, a 501B Open-Weight Model, but Has Not Named a Licence
A.I. / news
Reflection Announces Beam, a 501B Open-Weight Model, but Has Not Named a Licence
The weights are promised for later in October. Reflection's own post gives scores and training scale, and says nothing on licence terms or API pricing.

Reflection AI announced Beam on Monday, a 501-billion-parameter mixture-of-experts model with 23 billion active parameters, and said it will release the weights later in October. The company's launch post does not state a licence.
As of Tuesday, access is by waitlist at platform.reflection.ai. The post lists no API price. The weights, a technical report, a model card and developer artifacts are all described as coming "later this month".
What Reflection says Beam is
Beam was pretrained on 23.8 trillion tokens drawn from the web and what the post calls proprietary licensed datasets. Its context window is 256,000 tokens during reinforcement learning, extended to 1 million tokens of effective length.
Pretraining ran on 6,144 Nvidia GB300 NVL72 GPUs in under four weeks, with a goodput of 92.3%. The reinforcement-learning run used 10,500 GB300 GPUs for four weeks and produced more than 100 million rollouts across roughly 1.3 billion sandboxes.

SiliconANGLE reported that Reflection raised money at a $25 billion valuation a few months ago, and that it signed a $6.3 billion deal with SpaceX to rent GB300 NVL72 systems. Misha Laskin and Ioannis Antonoglou are named there as its founders.
The scores are Reflection's own
Every figure below comes from the launch post. No independent party is named as having run them.
| Benchmark | Beam score |
|---|---|
| SWE Bench Verified | 80.9 |
| Terminal Bench v2.1 | 80.1 |
| SWE Bench Pro v2-Hard | 77.2 |
| SWE Bench Pro v1 | 65.5 |
| GPQA Diamond | 90.5 |
| AIME 2026 | 97.8 |
- SWE Bench Verified80.9 score
- Terminal Bench v2.180.1 score
- SWE Bench Pro v2-Hard77.2 score
- SWE Bench Pro v165.5 score
Source: Reflection AI launch post, accessed 2026-10-06
Reflection calls Beam "competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks". It says Kimi K3 remains ahead on raw capability, and that Beam's edge is efficiency: on advanced reasoning benchmarks it matches GLM-5.2 while using 3 to 4 times less inference compute.
In one test, Reflection recreated a viral land-and-water puzzle on a 180 by 90 grid of 16,200 points. Beam got 95.5% of points right, which the post places between Opus 5 at 92.5% and Fable 5 at 97.8%.
What the post leaves out
The post names no licence. Open-weight releases differ on whether commercial use, fine-tuning for resale or redistribution is allowed, and Reflection's own marketing calls Beam "open-weight" without saying which terms apply. SiliconANGLE headlined it open source, a label that needs a licence to check.
The post also gives no download size and no minimum hardware. At 501 billion parameters, the weights alone will not fit on a single consumer GPU, but Reflection has not said what precision it will ship.
Efficiency claims such as the 3 to 4 times figure depend on the benchmarks and token budgets Reflection chose, and a technical report is the earliest place to check them.
Reflection said the weights, report and model card will be published by the end of October. The licence text, once the weights are up, decides whether Beam belongs alongside Aleph Alpha's Apache-licensed Kolibri-1 or Cloudflare's Clef models.
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.