Reflection introduced Beam, which it calls its first open-weight model, in a blog post dated October 5, 2026. The post says Beam is “a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active”, and that on advanced reasoning benchmarks it “achieves scores comparable to GLM-5.2 while using 3–4× less inference compute”. The post also says the weights are not out yet: “We will release the weights, technical report, model card, and developer artifacts later this month”.
Everything below is Reflection’s own description, from its Introducing Beam post. The benchmark scores are Reflection’s, and this publication did not run Beam or any other model.
Business Pill · HOW WE KEEP SCORE
A one-minute explainer of the idea behind this story: a proxy measure. It teaches the concept, not this story’s figures.
The key insight: The post’s efficiency claim rests on an estimate that it labels approximate: a formula that multiplies active parameters by generated tokens and leaves out prompt prefill, context-dependent attention operations and serving overhead. This publication does not say that the 3–4× figure is a measured cost, because the post says it is not.
What Reflection Says Beam Is
The post calls Beam “Reflection’s first open-weight model”, “built for coding, reasoning, and agentic workloads”. It adds: “Although Beam is text-only, it can work with information from other modalities when represented as text”.
On status the post says: “Beam is undergoing final red-teaming and evaluations”. It says Reflection is “making this early version of Beam available to a select group of users”, and it links a waitlist for early access.

Update, 16 October: Beam’s weights are out
The release this piece described in the future tense has happened. In a blog post dated October 16, 2026, Reflection says it has “released Beam’s model weights on Hugging Face under an Apache 2.0 license”, together with a model card with final benchmarks and a Beam Safety Technical Report. It is launching with what it calls “an ecosystem of over 40 partners and integrations”.
Per Hugging Face’s launch post written with the Reflection team: 501B total parameters with 23B active per token, a 1M-token context window, text in and text out, and checkpoints in BF16, FP8 and NVFP4 plus a speculative-decoding drafter. The BF16 checkpoint is about 1 TB; NVFP4 brings it to about 280 GB, and Hugging Face suggests planning for four B200 GPUs. Reflection is also releasing Mirror CLI, the open-source harness it says it built and trained Beam on.
The benchmark table Hugging Face publishes is “as provided by the Reflection team”. In BF16 it lists SWE-bench Verified at 81.6%, Terminal Bench v2.1 at 80.1% and GPQA Diamond at 90.6%. These are the company’s own numbers, not independent results.
What Was Posted on X
Reflection’s own post on 16 October announcing that Beam is available with open weights under the Apache 2.0 license, listing the weights, tokenizer, reference implementation, BF16/FP8/NVFP4 checkpoints, a drafter model and its Mirror CLI harness, with more than 40 partners.
Databricks congratulated Reflection on the Beam launch on 16 October, quoting Reflection’s post that Beam is available across platforms with open weights under Apache 2.0 and an ecosystem of 40+ partners.
What Is Released and What Is Promised
The post uses the future tense for the release. It says “This month, we will release the weights under an Apache 2.0 license, along with documentation and the full stack for running, evaluating, and fine-tuning the model”, and “We will be launching Beam with an ecosystem of distribution partners, as well as integration with a broad range of open source libraries and harnesses”.
This publication found no more specific date than “later this month” in the post. On safety it says “We will publish the results of our safety evaluations in our model technical report”.
The Efficiency Claim
The claim sits in one sentence: “On advanced reasoning benchmarks, it achieves scores comparable to GLM-5.2 while using 3–4× less inference compute”. The next sentence adds: “Efficiency gains are even more pronounced when comparing to models in the 2T+ parameter family like Qwen 3.8-Max, which require significantly more inference compute per token”.
The post says how the compute was counted. The note under its Figure 2 says Reflection estimated “generation forward-pass compute as FLOPs ≈ 2 × active parameter count × mean generated tokens per attempt”, and that “Generated tokens include both reasoning and the final answer”. It says the estimates “exclude prompt prefill, context-dependent attention operations, and serving overhead, so they represent an approximate compute comparison rather than measured inference cost”.
For the other models, the note says “We use Artificial Analysis and DataCurve as sources for other model’s evals”. For mixture-of-experts models it says “we used the parameters activated per token rather than the total model size”.
On raw capability the post is direct about one model: “Where frontier open models like Kimi K3 remain ahead on raw capability, Beam’s advantage is efficiency at inference time”. It says Beam is “competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks”.
What the Benchmark Table Shows
The post’s table lists seven benchmarks. For three of them it gives a score for both Beam and GLM 5.2. On DeepSWE v1.1 it lists Beam at 44.4 and GLM 5.2 at 44.0. On SWE Bench Pro v1 it lists 65.5 and 62.1. On Terminal Bench v2.1 it lists 80.1 and 81.0.
On DeepSWE v1.1 and Terminal Bench v2.1 the table lists higher scores for four other models. On Terminal Bench v2.1 it lists GLM 5.3 at 88.2, Kimi K3 at 88.3, Qwen 3.8 Max at 86.6 and DeepSeek V4.1 Flash at 90.6, against Beam’s 80.1. On DeepSWE v1.1 it lists GLM 5.3 at 61.0, Kimi K3 at 68.0, Qwen 3.8 Max at 51.0 and DeepSeek V4.1 Flash at 74.2, against Beam’s 44.4.
On SWEBench Verified the table gives Beam 80.9 and lists scores only for Inkling, at 77.6, and Nemotron 3 Ultra, at 70.7. The post’s note says “NR denotes scores that have not been reported”.
How Reflection Says It Trained Beam
On pretraining, the post says Beam was “pretrained end-to-end in under four weeks on a cluster of 6,144 NVIDIA GB300 NVL72 GPUs” on “23.8 trillion diverse high-quality tokens”. It says midtraining “extends Beam’s effective context length to 1M tokens”.
On reinforcement learning, it says the run “generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training”, with “a maximum context length of 256K tokens” and “nearly one million environments”. Reflection writes: “We believe this is one of the largest scale RL runs conducted by any open lab to date”.
The post also reports the pretraining run’s reliability: “nine semi-automatic rewinds” and a goodput that “reached 92.3% towards the end”.
Rewarding Short Answers
On token use, the post says Reflection trained Beam “with a controllable length penalty that rewards successful solutions while discouraging unnecessary tokens”. It says users can set a “reasoning effort parameter”: “lower settings favor shorter responses, while higher settings allow longer reasoning to improve performance on demanding tasks”.
The Structural Read
The post builds its efficiency case on two things it says about the model: a sparse Mixture-of-Experts design with 501 billion total parameters, of which 23 billion are active, and a length penalty that rewards successful solutions while discouraging unnecessary tokens. Its compute formula multiplies the active parameter count by the mean generated tokens, so both choices enter the estimate.
The table in the post lists near-equal scores for Beam and GLM 5.2 on the three benchmarks where it gives both. On the two benchmarks this article compared across the table, it lists higher scores for four other named models. The post itself says Kimi K3 remains ahead on raw capability and describes Beam’s advantage as efficiency at inference time.
The release is described in the future tense. The post says the weights under Apache 2.0, the technical report, the model card and developer artifacts come later this month, with early access through a waitlist. The scores and the compute figure are Reflection’s own, so none can yet be checked against released weights.
Reflection, Introducing Beam
“These estimates exclude prompt prefill, context-dependent attention operations, and serving overhead, so they represent an approximate compute comparison rather than measured inference cost.”
Three Implications
THE 3–4× FIGURE IS AN ESTIMATE The post’s note says the figure counts active parameters times mean generated tokens and excludes prompt prefill, context-dependent attention operations and serving overhead. It calls the result an approximate compute comparison rather than measured inference cost.
THE SCORES ARE REFLECTION’S OWN The post says it took other models’ evaluations from Artificial Analysis and DataCurve, and the table marks scores not reported as NR. This publication did not reproduce any score.
THE WEIGHTS ARE PROMISED, NOT RELEASED The post says the weights, technical report and model card come later this month. This publication found no more specific date in the text it read, and it did not read a technical report or model card.
What Is Not Established
The benchmark scores and the 3–4× figure are Reflection’s own. The post says it took other models’ evaluations from Artificial Analysis and DataCurve, and it describes its compute figure as an approximation. This publication did not run Beam, did not reproduce any score, and did not read a technical report or model card, because the post says they are not out yet.
This publication read the text of the Introducing Beam post and its benchmark table. It did not read the data inside the post’s charts and animations, did not read the early-access terms, and did not contact Reflection. The text it read does not say whose cluster the GB300 GPUs sit in, and this publication does not say. Its separate report on Reflection’s compute deal with SpaceX is a different story, and nothing in the Beam post links the two.
No score here has been checked against the weights, because the weights have not been released. The post says they will be.
The Bottom Line
Reflection’s October 5 post introduces Beam as a 501-billion-parameter open-weight model with 23 billion active parameters. It says Beam scores comparably to GLM-5.2 on advanced reasoning benchmarks with 3–4× less inference compute, a figure the post itself calls an approximate comparison rather than a measured cost. It says the weights under Apache 2.0, the technical report and the model card will come later this month. Until then, the scores are Reflection’s own, and nothing in this article has been tested.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We drew this piece entirely from Reflection’s Introducing Beam blog post, which we read on 6 October 2026. We did not run Beam, did not consult a technical report or model card (Reflection says neither is out yet), and did not read the data inside the post’s figures. We did not contact Reflection, and we have not checked their claims independently. The benchmark scores and the 3–4× figure are Reflection’s own. Nothing here constitutes a forecast or any form of advice.
Sources: Reflection, ‘Introducing Beam’, October 5, 2026, full text as supplied)









