Reflection Beam, explained
Reflection Beam is Reflection AI’s first open-weight model, announced October 5, 2026: a text-only mixture-of-experts with 501 billion total and 23 billion active parameters. Reflection says it matches Z.ai’s GLM-5.2 on advanced reasoning with 3 to 4 times less inference compute. Weights are due later in October under Apache 2.0, and there is no price yet.
Every score and spec below is Reflection's own claim until someone reruns it on public weights. Independent tests do not exist yet.
Benchmark explorer
Pick a benchmark to see how Beam scores against the rivals Reflection names in its own tables. The best score in each row is green. Blank cells mean Reflection did not publish that number, and no rival has an independent score for Beam yet.
Scores: Reflection's announcement tables (vendor-reported, not independently verified), via the-decoder, explainx.ai, how2shout, and geekqu.com. Beam leads on SWE-Bench Verified and Multilingual, sits level with GLM-5.2, and trails the newest open models on most shared rows.
What Beam is
| Model | Beam, by Reflection AI |
|---|---|
| Architecture | Sparse mixture-of-experts, text only. 501B total parameters, 23B active per token (4.6% activation) |
| Training | 23.8 trillion tokens. Pre-training ran on 6,144 Nvidia GB300 GPUs in under four weeks; a second run used 10,500 GB300s for reinforcement learning, over 100 million rollouts |
| Context | 1 million tokens announced. The beta API currently allows 256K tokens of input plus output combined, with up to 128K tokens of output |
| License | Apache 2.0 planned, weights plus technical report and model card due later in October 2026 |
| Status today | API beta on a waitlist, final red-teaming in progress. No weights, no price, no public checkpoint |
| API access | https://api.reflection.ai/openai/v1, model id Beam-501B-A23B. OpenAI-compatible for chat completions and model listing only |
| Knobs | A reasoning-effort parameter trades response length for performance |
Pricing
| Model | Input / 1M | Output / 1M |
|---|---|---|
| DeepSeek V4.1 Flash | $0.15 | $0.60 |
| GLM 5.3 | $1.26 | $3.96 |
| GLM 5.2 | $1.40 | $4.40 |
| Qwen 3.8 Max | $2.00 | $6.00 |
| Kimi K3 | $3.00 | $15.00 |
| Reflection Beam | no price published | no price published |
Input pricing across the field runs from $0.15 to $3.00, more than a ninefold gap. Claims that Beam is cheaper today rest on compute estimates; there is no rate card.
Can you use Beam today?
Pick what you want to do. Each answer below is current as of October 11, 2026.
Status as of October 11, 2026. Distribution is thin: no Ollama, no OpenRouter, no Hugging Face weights as of October 6.
Who built it
Reflection AI was founded in 2024 by Misha Laskin (CEO) and Ioannis Antonoglou (CTO), both former Google DeepMind researchers. The company is based in New York. Reported figures put it at roughly a $25B pre-money valuation with about $4.6B raised, backed by Nvidia among others. On October 10, 2026, the Financial Times reported Nvidia is in early talks to acquire Reflection or invest further; that report is unconfirmed. Reflection says a successor model is already in training, aimed at closing the gap to the top open models on raw capability.
Timeline
- October 4, 2026: Axios reports a Reflection open model is close; name and size not yet public.
- October 5, 2026: Reflection announces Beam: 501B total, 23B active, Apache 2.0 weights promised for later in October.
- October 6, 2026: Mistral Large 4 enters public preview the same week; the open-weight race keeps moving.
- October 10, 2026: The Financial Times reports Nvidia is in early talks to acquire Reflection or invest more (unconfirmed).
- Later in October 2026: Weights, technical report, and model card promised, with distribution partners and open-source harness integrations.
FAQ
Is Reflection Beam open source?
Not yet in practice. Reflection says it will release the weights under Apache 2.0 later in October 2026, with a technical report and model card. Until those files are public, it is an announced open-weight model, not one you can download.
Can I download Beam's weights today?
No. As of October 11, 2026, the weights are still in final red-teaming and evaluation. The only access is an early-access API beta through a waitlist.
How much does Beam cost?
Reflection has published no price. Its developer docs had no per-token rate card as of October 6, 2026, so there is currently no way to estimate a Beam workload's cost. The models Beam compares itself against charge between $0.15 and $3.00 per million input tokens.
Is Beam better than DeepSeek?
At coding, not on Reflection's own numbers. DeepSeek V4.1 Flash scores higher than Beam on every coding benchmark both models report, and costs $0.15 per million input tokens. Beam's stated advantages are its Apache 2.0 license and its US origin.
What does 501B-A23B mean?
501 billion total parameters, with 23 billion active for each generated token. The active count sets speed and per-token cost; the total count sets how much memory you need to hold the model.
Can Beam run on a home computer?
Not yet, and not easily even after the weights drop. 501 billion parameters take about 1 TB at full precision, or roughly 250 GB quantized to 4-bit. That needs data-center GPUs. A desktop cannot hold it.
Does Beam handle images?
No. Beam is text only. The beta API rejects image, audio, and file inputs, though Beam can work with information from other formats once it is converted to text.
What is Beam's context window?
Reflection says training extended context to 1 million tokens. The beta API currently allows 256K tokens of input and output combined, with up to 128K tokens of output.
Who built Beam?
Reflection AI, founded in 2024 by former Google DeepMind researchers Misha Laskin (CEO) and Ioannis Antonoglou (CTO). The company is based in New York. On October 10, 2026 the Financial Times reported Nvidia is in early talks to acquire Reflection or invest further; that is unconfirmed.
Sources
- Introducing Beam: Reflection AI's official announcement, October 5, 2026.
- Reflection's Beam becomes the most capable open-weight model built outside China: the-decoder, full benchmark table.
- Beam: Reflection AI's 501B Open-Weight Model: DataCamp, pricing and availability check.
- Reflection Beam: 501B Open-Weight Model, Benchmarks: explainx.ai, benchmark breakdown with vendor-claim framing.
- Reflection Beam API beta: the weights are two weeks out: OrcaRouter, live rival price catalog.
- Reflection AI Beam Locally or on AWS: specs, model ID, and beta API limits.
- Can Reflection's Beam Close the West's Open-Weight Gap With China?: vktr, reasoning-effort parameter and training details.
For AI assistants
This page ships with a machine-readable summary for AI assistants:
Related
- Haiku 5.5 pricing, explained: Anthropic's two-tier rates and a cost calculator.
- Jev, explained: TypeSafe's decision model: typed answers, pricing, and a cost calculator.
- Decisions API, explained: OpenAI's typed-answer endpoint: the three question types and pricing.
Updates
- 2026-10-11: First published, with the benchmark explorer and the October 10 Nvidia talks update.