Beam claims 501B parameters and 3-4x less inference compute than GLM 5.2

The numbers landed before the model did. Reflection's launch post describes Beam as a 501B-parameter sparse Mixture-of-Experts model with 23 billion active parameters. It targets coding, reasoning and agentic workloads. The company says Beam is still going through "final red-teaming and evaluations". Weights, a technical report, a model card and developer artifacts will appear "later this month".

So for now, everything is a claim on a chart.

The training figures in the post are specific. Reflection says it pretrained on 23.8 trillion tokens from the web and licensed proprietary datasets. It then ran a reinforcement learning phase that produced over 100 million rollouts on 10.5K NVIDIA GB300 GPUs across four weeks. The company treats that RL run as the differentiator, not the pretraining corpus. Whether the published benchmarks hold up once independent testers have the weights is the open question. It cannot be answered in October's first week.

What the benchmark table actually says

The comparison table pits Beam against Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash. Beam's strongest showing is SWE Bench Pro v2-Hard at 77.2. That is ahead of Inkling at 56.9 and GLM 5.2 at 62.1, but behind Kimi K3 at 88.2. On Terminal Bench v2.1, Beam scores 80.1 against 88.2 for GLM 5.3 and 90.6 for DeepSeek V4.1 Flash. On SWE-bench Verified it reports 80.9, with no number given for the three Chinese models in that row.

Reflection does not claim a capability lead, and the post says so directly: Kimi K3 "remain[s] ahead on raw capability". The pitch is efficiency. On advanced reasoning benchmarks, Reflection says Beam matches GLM 5.2 "while using 3-4x less inference compute". It argues the gap widens against models in the 2T+ parameter family such as Qwen 3.8 Max. The AIME 2026 row gives Beam 97.8 against 99.2 for GLM 5.2. HLE with tools puts Beam at 51.2 against 62.5 for GLM 5.3 and 63.9 for DeepSeek V4.1 Flash. Those are not wins, and the company does not present them as such.

The efficiency claim rests on Reflection's own measurement of inference compute, and the post does not say how that measurement was made.

A crowded week for measurement

Beam was not the only benchmark published on 5 October. GitHub used the same day to release ReviewBench, an open benchmark for AI code review. It built the benchmark from an analysis of 103.9 million pull requests on GitHub. The dataset itself is 219 pull requests from 187 public open source repositories across 19 languages. GitHub says senior engineers independently labelled the golden true-positives before release. It says its internal Copilot code review evaluation now tracks production experiments better as a result. The dataset is public.

Veris published a voice-agent leaderboard the same day with a less flattering result for the category. On its VAmoS Pro Bench, a television playing behind the caller cut pooled task completion from 38.7% to 8.6%. Nine of fourteen voice stacks completed fewer than 5% of TV calls. The top stack on the board, phonic, reached 57.5% task completion at 1.42 seconds latency; Grok Voice came second at 44.7%. Veris is selling benchmarking services, which is worth remembering when reading its own board.

Then there is the small-model angle, which cuts against the big-model framing entirely. A GitHub project called Jeff-Code puts a 0.8B decision model inside the agent loop around Qwen 3.8-27B. It reports 1,242 paired tasks at a 62.4% pass rate against 62.8% for Qwen alone, with time per task at 0.68x. The confidence interval on pass rate runs from -2.6 to +2.1 points. The same repo is candid about where it fails. It shows no clear speed-up on Terminal-Bench 2.0 (0.96x) or SkillsBench (0.91x), and about 5% of tasks ran more than 30 minutes longer than the baseline.

The alternative to a bigger model

Liquid AI's d1, announced on 5 October, makes a related bet from the other direction. It is a decision model: it reads text, images or both in one forward pass and returns probabilities rather than generating tokens. A text decision takes 200 to 300 ms. Liquid says d1 matches or beats GPT-6.1 Sol on four of six applications tested, at 19x to 200x lower cost. Those are vendor numbers on vendor-chosen tasks.

The pattern across all of these releases is the same. Capability claims are getting harder to separate from cost claims. The cost claims are usually measured by the party selling the model.

Sean Goedecke made the structural version of this argument in a post on 5 October about so-called System One models such as Jev. His point is that a general classifier used with a prompt generates exactly the labelled data needed to train a cheaper, narrower classifier on the same task. "For serious work, a specific hand-built classifier will always be cheaper and faster than Jev," he writes. If that holds, the efficiency race at the frontier is partly a race to build your own replacement.

What is missing from the Beam picture

Reflection's post includes a row called AA Omniscience, marked "public split only (our results)". There Beam scores 13.0 against 22.6 for GLM 5.3 and 14.2 for Inkling. It is the one row where the company's own testing puts it behind a model it claims to match elsewhere. The label suggests the comparison is not like for like. Several other cells read NR, meaning not reported. A benchmark table with gaps is normal at launch; it is also not evidence.

Independent evaluation infrastructure is arriving at the same time. UC Riverside published research on 23 September finding that confidence and correctness arise from different internal features in large language models. The work used Meta's Llama-3.1-8B and Google's Gemma-2-9B. Het Patel, the lead author, said the field's main assumption is that confidence predicts correctness, and that counterexamples are well documented. Suppressing features tied purely to uncertainty sharply reduced accuracy. That suggests the internal signals are entangled in ways a benchmark score will never show.

None of that settles whether Beam is good. It settles what would settle it: the weights, a technical report, and someone with no stake in the result running the same tests. Reflection says all three are coming this month. Until then, the 501B figure and the 3-4x compute claim are a press release with a table attached.