
What Reflection announced
On October 5, 2026, Reflection AI — the Brooklyn startup founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou and backed by Nvidia — unveiled Beam, its first open-weight model. In its launch blog post "Introducing Beam," the company describes it as a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning, and agentic workloads.
Beam is text-only: it handles text input and output, and can work with other modalities only when they are represented as text. Reflection says weights, a technical report, model card, and developer artifacts will ship under the Apache 2.0 license later in October 2026; until then, early access runs through a waitlist on the company's platform while the model finishes final red-teaming and evaluations.
The spec sheet
- 501 billion total parameters, 23 billion active per token (sparse MoE)
- 1 million-token effective context window
- Reasoning effort control — low, medium, high, xhigh, max — trading response length against reasoning depth
- Pretrained on 23.8 trillion curated tokens
- RL post-training: 100M+ rollouts on 10.5K Nvidia GB300 GPUs over four weeks, ~1 million coding/agentic/STEM environments, ~1.3 billion sandboxes
- Apache 2.0 weights promised later in October 2026
- Early access via waitlist at platform.reflection.ai
The benchmark claims — and the caveats
Reflection positions Beam against Chinese open models: it says Beam "advances the Western open-weight frontier," is competitive with Z.ai's GLM-5.2 and approaching Alibaba's Qwen 3.8-Max on coding and agentic tasks, while Kimi K3 remains ahead on raw capability. On Beam's own table: DeepSWE v1.1 44.4 (GLM-5.2: 44.0), Terminal Bench v2.1 80.1 (GLM-5.2: 81.0), SWE-Bench Pro v1 65.5 (GLM-5.2: 62.1, Thinking Machines' Inkling: 54.3), and SWE-Bench Verified 80.9 (Inkling: 77.6).
Why the pitch is efficiency, not raw size
Beam's headline is inference efficiency: Reflection claims comparable advanced-reasoning scores to GLM-5.2 while using 3-4x less inference compute, and an even bigger edge over 2T+ parameter models like Qwen 3.8-Max. The mechanism is a controllable length penalty in RL training that rewards solving tasks with fewer tokens, plus a reasoning-effort knob so users can match compute to the task. For enterprises, that translates into "more intelligence per token" — lower serving bills on agentic coding workloads.
How Beam compares to the open-model field
| Model | Parameters | Focus | License status (Oct 2026) |
|---|---|---|---|
| Beam (Reflection) | 501B total / 23B active MoE | Coding, reasoning, agents | Apache 2.0 promised late Oct 2026 |
| GLM-5.2 (Z.ai) | ~744B total / 40B active MoE | Open coding/reasoning | Already released |
| Kimi K3 (Moonshot AI) | Not disclosed | Open reasoning | Already released — leads Beam on raw capability in Reflection's own table |
| Inkling (Thinking Machines) | Not disclosed | Multimodal open model | Already released — multimodal, where Beam is text-only |
What to do now
Beam is a launch announcement plus a schedule, not a downloadable checkpoint. If you run coding agents, the practical moves are: join the waitlist for early API access, prepare an evaluation suite on your own repos (DeepSWE-style and Terminal-Bench-style tasks), and watch for the Apache 2.0 weight drop later in October 2026 — that is the moment independent scores from trackers like LMArena and Artificial Analysis can confirm or revise Reflection's claims.
“Beam advances the Western open-weight frontier and is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks.”
Frequently Asked Questions
What is Reflection AI Beam?
Beam is Reflection AI's first open-weight model, unveiled October 5, 2026: a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, built for coding, reasoning, and agentic workloads. It is text-only, with a 1 million-token context window and a user-controllable reasoning-effort setting.
When will Beam's weights be released?
Reflection says the weights, technical report, model card, and developer artifacts will ship under the Apache 2.0 license later in October 2026. Until then, the model is in final red-teaming and evaluations, with early access through a waitlist.
How does Beam compare to GLM-5.2 and Kimi K3?
Reflection claims Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute, but says Moonshot AI's Kimi K3 remains ahead on raw capability. Beam also outscores Thinking Machines' Inkling on four coding tests in Reflection's table. All scores are company-reported and unverified.
Is Beam's 3-4x efficiency claim proven?
Not independently. Reflection's own blog calls the comparison approximate — it estimates forward-pass FLOPs from active parameters and generated tokens, excluding prompt prefill, attention operations, and serving overhead. Independent verification will only be possible after the weights are public.
Who is Reflection AI?
A Brooklyn startup founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, backed by Nvidia. Reflection has signed large compute deals, including capacity at SpaceX's Colossus 2 data center, to train its models on Nvidia GB300 GPUs.







