| Entity | Reflection |
|---|---|
| Date | 5 Oct 2026 |
| Region | U.S. & Allies |
On October 5, 2026, Reflection introduced Beam, its first open-weight model: a sparse Mixture-of-Experts with 501 billion total parameters and 23 billion active, built for coding, reasoning, and agentic workloads. Weights are not out. The post says Beam is in final red-teaming and evaluations, with early access by waitlist, and that the weights, technical report, model card, and developer artifacts will be released later this month under an Apache 2.0 license.
Who this affects
- For exporters: The post commits to releasing the weights this month under Apache 2.0, with documentation and the stack for running, evaluating, and fine-tuning; until that release, access is a waitlist for a select group.
- For investors: Reflection reports Beam competitive with larger open models such as GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks, with scores comparable to GLM-5.2 on advanced reasoning benchmarks at 3–4× less inference compute; it states Kimi K3 remains ahead on raw capability.
- For counsel/compliance: The Apache 2.0 release is prospective. The October 5 post says final red-teaming and evaluations are still underway, and the early version is limited to a select group via waitlist at platform.reflection.ai.
- For policymakers: The post frames Beam as advancing the Western open-weight frontier against named comparators GLM 5.2, Qwen 3.8-Max, and Kimi K3, on a text-only model.
The read
The event is an introduction plus a release commitment, not a weight drop. The efficiency claim is the load-bearing commercial point, and the post qualifies it: the Figure 2 comparison estimates generation forward-pass compute as FLOPs ≈ 2 × active parameter count × mean generated tokens per attempt, excludes prompt prefill, context-dependent attention, and serving overhead, and is described as an approximate compute comparison rather than measured inference cost. Other models’ evals in that figure are sourced by Reflection to Artificial Analysis and DataCurve. The open question the record can already bound is timing: Apache 2.0 weights, the technical report, and the model card are stated for later this month, after red-teaming.
What was the action?
Reflection introduced Beam on October 5, 2026, as its first open-weight model. Beam is a sparse Mixture-of-Experts with 501 billion total parameters and 23 billion active, pretrained on 23.8 trillion tokens from the web and proprietary licensed datasets. A high-compute reinforcement-learning run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks, with a maximum context length of 256K tokens. Training and grading used approximately 1.3 billion sandboxes. The model is text-only; the post says it can use other modalities only when represented as text.
What’s the penalty / scope, and how is it set?
No penalty. The action is a model introduction with a stated release scope. Reflection says it will release the weights later this month under an Apache 2.0 license, together with the technical report, model card, developer artifacts, documentation, and the stack for running, evaluating, and fine-tuning. Reported scores in its table include DeepSWE v1.1 at 44.4, SWE-bench Pro v1 at 65.5, Terminal Bench v2.1 at 80.1, and SWE-bench Verified at 80.9. On advanced reasoning benchmarks it reports scores comparable to GLM-5.2 at 3–4× less inference compute, with larger efficiency gaps against 2T+ models such as Qwen 3.8-Max. Users can set a reasoning-effort parameter that trades response length against performance.
Who else is exposed?
Named in Reflection’s comparison table: Inkling, Nemotron 3 Ultra, GLM 5.2, GLM 5.3, Kimi K3, Qwen 3.8 Max, and DeepSeek V4.1 Flash. The post positions Beam for enterprise coding and agentic workloads. Training hardware named in the post is NVIDIA GB300. Counterparties beyond the waitlist are not specified in the record.
What happens next
Final red-teaming and evaluations are still underway. The stated next release is the weights under Apache 2.0, plus the technical report, model card, and developer artifacts, later this month. Early access remains a waitlist. Whether the weights ship on that timetable is the open item in the post; the October 5 text does not give a calendar date.