Reflection AI has made its case for Beam more specific, publishing scores that put the model near Z.ai’s GLM-5.2 on two benchmarks. The company also says Beam is approaching Alibaba’s Qwen 3.8-Max on selected coding and agentic tasks, but its own table shows Qwen, Kimi K3 and DeepSeek V4.1 Flash ahead of Beam on both listed tests. A South China Morning Post report on Tuesday framed the launch as a challenge to Chinese leadership in open-weight models.
The comparison is Reflection’s, not an independent evaluation. And Beam’s weights are still not publicly downloadable: Reflection offers early access by sign-up and says the weights and developer materials are planned for later in October, without a release date.
Our earlier report on Beam covered the announcement and the wait for public weights. The new scorecard gives readers more to assess, but outside developers cannot yet run the model from public weights to check the results for themselves.
How does Beam compare with Chinese models?
On Terminal Bench v2.1, Reflection’s table gives Beam a score of 80.1, compared with 81.0 for GLM 5.2, 86.6 for Qwen 3.8 Max, 88.3 for Kimi K3 and 90.6 for DeepSeek V4.1 Flash. On DeepSWE v1.1, the scores are 44.4 for Beam, 44.0 for GLM 5.2, 51.0 for Qwen 3.8 Max, 68.0 for Kimi K3 and 74.2 for DeepSeek V4.1 Flash, according to Reflection’s benchmark table.
That makes the comparison with GLM close on these two tests: Beam is slightly behind on Terminal Bench and slightly ahead on DeepSWE. It falls behind Qwen, Kimi and DeepSeek on both. Reflection describes Beam as competitive with GLM-5.2 and approaching Qwen 3.8-Max on selected coding and agentic tasks; the company also acknowledges that Kimi K3 remains ahead on raw capability.
The results support a limited comparison, not a claim that Beam is better than Chinese models overall. They are scores from Reflection’s own evaluation, and the supplied reporting does not establish that independent evaluators have reproduced them. Two benchmarks also cannot show how the models compare across other uses or in deployed systems.
What does Reflection say about Beam’s efficiency?
Reflection describes Beam as its first open-weight model, built for coding, reasoning and agentic workloads. The company says it is a mixture-of-experts model with 501 billion total parameters and 23 billion active parameters.
Reflection says Beam was pretrained on 23.8 trillion tokens from web and proprietary licensed data. The company also reports that its reinforcement-learning run involved more than 100 million rollouts on 10,500 NVIDIA GB300 GPUs over four weeks. It says training and grading used about 1.3 billion sandboxes. These are company-reported details about the model’s development, not independent evidence of its performance.
The company estimates that Beam uses three to four times less inference compute than GLM-5.2 on advanced reasoning benchmarks. Reflection says the estimate covers generation compute but excludes prompt prefill, context-dependent attention operations and serving overhead.
Those exclusions limit what the comparison can establish for organizations considering operating costs. The estimate is not a measured comparison of serving costs, and it does not show that Beam will be cheaper, faster or less energy-intensive to run in deployment. A comparable test of real serving conditions would be needed to assess those outcomes.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
When can developers use Beam’s weights?
For developers who want to run or adapt Beam independently, the public weights and accompanying materials are the practical test of the open-weight offer. A sign-up for early access is not the same as public access to the weights.
Reflection says Beam is undergoing final red-teaming and evaluations. It plans to publish the weights, a technical report, a model card and developer materials later in October, but has not given a specific date. Those materials could give developers more information about the model and a basis for independent testing once the weights are available.
Until then, readers can examine the company’s published scorecard, but they cannot independently run the model from public weights. The terms governing the eventual release are also among the details developers will need to assess.
Does Beam change the open-weight model landscape?
Semafor has described Chinese companies as leaders in the open-model field and reported that some Western companies use Chinese models for lower-cost routine tasks. Reflection is pitching Beam as another option for businesses and governments seeking to build AI systems without relying on Chinese models.
Misha Laskin, Reflection’s co-founder and CEO, told Semafor: “They don’t really have very good options today.” That is Laskin’s view of the market opportunity, not independent evidence of demand or proof that Beam meets it.
Beam adds a U.S.-based entrant to a field with strong Chinese competition, but one unreproduced scorecard does not establish a shift in that balance. The next test is whether Reflection publishes the weights and documentation as planned later in October, and whether outside evaluators can reproduce the benchmark results and examine the compute estimate under comparable conditions.