Reflection's Beam: 501B Open-Weight Model, No Weights Yet, Self-Reported Benchmarks

The blog post landed on 5 October, and it is careful about one thing: Beam is not finished. "Beam is undergoing final red-teaming and evaluations," the company writes, and the weights, technical report, model card and developer artifacts are promised "later this month." Early access runs through a signup form.

What Reflection has published is a specification sheet and a table of benchmark scores it ran itself. That distinction matters for anyone planning to evaluate the model on their own workloads. Nothing in the post says when independent testing will exist.

The architecture is a sparse Mixture-of-Experts: 501 billion total parameters with 23 billion active per token, pretrained on 23.8 trillion tokens that Reflection describes as web data plus proprietary licensed datasets. The reinforcement learning stage ran over 100 million rollouts on 10.5K NVIDIA GB300 GPUs across four weeks of training. Those are the company's own figures. They are also the kind of number an outside party finds hard to check.

The efficiency claim is the pitch

Reflection frames Beam as a workhorse rather than a frontier leader. On its own table, Beam scores 80.9 on SWE-bench Verified against 77.6 for Inkling and 70.7 for Nemotron 3 Ultra. On Terminal Bench v2.1 it reports 80.1, behind GLM 5.3 at 88.2, Kimi K3 at 88.3 and Qwen 3.8 Max at 86.6.

The company concedes the capability gap in plain language. "Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time," the post says. It claims scores comparable to GLM 5.2 on advanced reasoning benchmarks while using 3 to 4 times less inference compute, with the gap wider against models in the 2T+ parameter family such as Qwen 3.8 Max.

If that holds up, it changes the maths for teams running coding agents at volume. Inference cost, not benchmark rank, decides whether an agent runs on every pull request or only on the ones a human flags. It also puts pressure on Chinese open-weight labs, which have led on price per token. What remains unknown is the licence, the weight format, and whether the efficiency claim survives a third-party harness. None of that is in the post.

Benchmarking is having a week of its own

The timing is awkward in a useful way. On the same day, GitHub published ReviewBench, an open benchmark for AI code review built from an analysis of 103.9 million pull requests on the platform. The dataset contains 219 pull requests from 187 public open source repositories across 19 languages, with language and repository size distributions matched to GitHub overall. GitHub says senior engineers independently labelled the golden true positives before release, and that it deliberately weighted pull request size toward "the reviewable middle and tail" rather than mirroring the platform exactly.

GitHub also says the benchmark improved its ability to predict whether changes to Copilot code review would help in production. That is the interesting part: a benchmark used internally as a filter before shipping, not as a leaderboard.

Smaller teams are making the same argument from the other direction. A developer release for Jeff-Code, dated 5 October, puts a 0.8B decision model inside the Pi coding agent loop alongside Qwen 3.8-27B. Across 1,242 paired tasks it reports a pass rate of 62.4% against 62.8% for Qwen alone, a paired difference of minus 0.2 points with a 95% interval of minus 2.6 to plus 2.1, while average time per task fell to 0.68 times the baseline. The project is explicit that the speed-up is not uniform: it reports 0.96 times on Terminal-Bench 2.0 and 0.91 times on SkillsBench, meaning no clear gain on either.

That kind of disclosure is rarer than it should be. It also makes the aggregate claim easier to trust than a single headline percentage.

Decision models, vision and the small-model push

Liquid AI used 5 October to launch d1, which it calls its first decision model and the first to handle images as well as text. Decision models return probabilities for a set of questions in one forward pass instead of generating tokens. Liquid says a text decision takes 200 to 300 ms, that d1 matches or beats GPT-6.1 Sol on four of six test applications, and that it costs 19 to 200 times less than both that model and Claude Opus 5.5.

The company reports 85 to 97% accuracy sorting good and defective parts in an industrial inspection demo on the public VisA dataset, and says the model was never trained for that task. It also reports removing 52% of tokens in a coding agent context compaction test while keeping every output the task needs. Those are vendor numbers, run by the vendor, and should be read that way.

The same caution applies to Interfaze 1.1 Lite, an open-weight mixture-of-architectures model posted to Hugging Face on 5 October. It pairs a reasoning core with specialist models for documents, speech, segmentation and forecasting, and the authors say the whole thing runs on one 80 GB GPU with no external services. Its published comparison table puts it at 73.2 on MMMU-Pro against 71.1 for Interfaze and 40.4 for GPT-5.4-Mini, and at 3.01 word error rate on VoxPopuli-Cleaned against 2.4 for Interfaze. Mixed results on the vendor's own grid.

Elsewhere in the same cluster, Veris published a voice-agent leaderboard on 5 October built from 2,694 simulated utility billing calls. It found that a television playing behind the caller cut pooled task completion from 38.7% to 8.6%, and that nine of fourteen voice stacks completed fewer than 5% of those calls. The top entry, Phonic, completed 57.5% of tasks at 1.42 seconds latency. One exam, one utility's policies, as Veris itself notes.

Why the missing weights matter

Reflection's claim is the most consequential of the batch because it is about cost at scale. It is also the one nobody outside the company can test yet. The company has said the artifacts arrive this month. Until then, Beam is a promise with a parameter count attached.

The rest of the day's releases point the same way: smaller models, specialist models, and harnesses that decide when a large model needs to think at all. That is a cheaper direction than another trillion parameters. It is also harder to benchmark, which is why the open datasets published alongside them may end up mattering more than the models.