Reflection AI unveils Beam: 501B-parameter open-weight MoE with 1M-token context
Reflection AI has unveiled Beam, a text-only mixture-of-experts model with 501 billion total parameters, 23 billion active, and a 1 million token context window. The company claims it matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. Weights and the full technical report are due later this month.
How Beam ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| AA Intelligence Index | 17.7 index | 44 of 56 | Artificial Analysis, Nov 2026 ↗ |
| Speed (tok/s) | 258 tok/s | 7 of 105 | Artificial Analysis, Nov 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Reflection AI has unveiled Beam, a text-only mixture-of-experts (MoE) model with 501 billion total parameters, 23 billion active parameters, and a 1 million token context window. The company announced it Monday, October 5, 2026. According to Reflection, Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using "3-4x less inference compute." Those performance claims have not been independently verified.
Specifications
| Spec | Beam |
|---|---|
| Architecture | Mixture-of-experts, text-only |
| Total parameters | 501B |
| Active parameters | 23B |
| Pre-training data | 23.8T tokens |
| Context window | 1,000,000 tokens |
| Post-training | High-compute reinforcement learning |
| Pricing | Not yet disclosed |
| Training cutoff | Not yet disclosed |
| Weights | Due this month, per Reflection |
Reflection describes Beam as a "workhorse model" for enterprises, the public sector, and developers. It says the model is trained for reasoning, coding, and agentic tasks at "a fraction of the token cost and inference time compute" of rivals.
For comparison, GLM-5.2 has roughly 744B total and 40B active parameters. Beam has about 33% fewer total parameters and about 42% fewer active parameters.
Performance claims
Reflection says Beam scores on par with GLM-5.2 on advanced reasoning benchmarks and outperforms today's leading Western open models. The company has not disclosed specific benchmark figures in the material available to us, and none have been independently confirmed.
Reflection's own results also show Beam outscoring Inkling, the open model from Thinking Machines Lab released in July, on four coding tests where both report scores. The comparison is not like-for-like: Inkling is multimodal, and Beam handles text only.
Availability
Reflection says it will release Beam's weights and full technical details this month. It also says the model will be distributed through hyperscalers and neoclouds, with integrations across open-source libraries at launch. Licensing terms have not been disclosed. Reflection did not respond to TechCrunch's requests for more information.
Company and compute
Reflection was founded in 2024 by two former Google DeepMind researchers and is based in Brooklyn. According to PitchBook, it has raised about $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners. Its last round valued the company at $25 billion pre-money.
This summer, Reflection signed deals collectively worth more than $7 billion with SpaceX and Nebius for access to Nvidia GB300 chips through 2029.
The company is targeting enterprises and sovereign customers with what it calls "AI factories." Institutions would train Reflection's models on their own proprietary data to build customized local systems. Reflection has begun testing a sovereign AI factory partnership with Shinsegae Group in South Korea. Axios reported that hedge funds and trading firms are also interested.
What this means
Beam is the strongest signal yet of a U.S. lab trying to compete with Chinese open-weight models on efficiency, not just capability. The headline claim is the 3-4x inference compute advantage. If it holds up, it matters more than any benchmark parity, because serving cost is the main variable for enterprises deploying open models at scale.
Three things need to be checked before the claims can be trusted. The first is the weights release, which will let outside teams reproduce the results. The second is the actual benchmark numbers, which Reflection has not yet put in front of independent evaluators. The third is the license, which determines whether sovereign and regulated buyers can use Beam the way Reflection's pitch implies.
The text-only design is a trade-off. It likely helps efficiency, but it leaves Beam without the multimodal capability that Inkling offers. The model's real test is whether it holds up on third-party reasoning and agentic evaluations once the weights are public.
Related Articles
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.
NVIDIA releases Nemotron-Labs-Sigma: 14 open 3B and 8B continuous diffusion language model checkpoints
NVIDIA released Nemotron-Labs-Sigma, a set of 14 open continuous diffusion language model checkpoints at 3B and 8B parameters. According to MarkTechPost, Sigma denoises text in a learned 16-D embedding space, distills to few-step samplers, and decodes up to 4.20x faster than autoregressive decoding.
Tencent releases Hy4 flagship weights: 770B-parameter MoE, 1,048,576-token context, conditional license
Tencent released downloadable weights for its flagship Hy4 on November 20, with a System Card. The multimodal MoE has a 770-billion-parameter language backbone, 49 billion active parameters per text token and a 1,048,576-token context window, under a license with commercial conditions for model-as-a-service operators.
Comments
Loading...