The announcement
On October 5, 2026, US startup Reflection AI introduced Beam, its first open-weight model: a sparse Mixture-of-Experts (MoE) with 501 billion total and 23 billion active parameters per token, built for coding, reasoning and agents. The company promises Apache 2.0 weights "later this month," with a technical report, model card and fine-tuning tooling.
Status as of Oct 12: the weights are not yet available. For now there's only a waitlisted beta API with no public pricing. The model is still in "final red-teaming and evaluations."
Spec sheet (per Reflection and coverage)
- Architecture: sparse MoE with routed experts, interleaved local/global attention, 52 layers (9to6ai). ID:
Beam-501B-A23B. - Text-only, June 2026 knowledge cutoff; the beta API lists 256K context.
- Pretraining: 23.8 trillion tokens on 6,144 NVIDIA GB300 GPUs in under 4 weeks.
- API: OpenAI-compatible endpoint (Chat Completions, streaming, tool calling, structured outputs). Responses, Embeddings, Batch and Files are not supported.
- Infrastructure: per SiliconANGLE, Reflection signed a $6.3B deal with SpaceX to rent GB300 NVL72 systems and raised at a $25B valuation.
Performance claims (unverified)
Reflection says Beam matches GLM-5.2 (Z.ai, ~250B more parameters) on reasoning tasks using 3–4x less inference compute, and approaches Qwen 3.8-Max. Reported figures include 80.9 on SWE-bench Verified.
These are the company's own numbers and, per Startup Fortune as cited by The Terminal, have not been independently verified.
Why it matters
- Competitive US open source: per SiliconANGLE, many top open models come from Chinese firms (Qwen, GLM, Kimi, DeepSeek). Beam would be the first from a US startup to show comparable performance.
- Efficiency: with 23B active, per-token cost is far below a dense model of that size, but you still need memory for all 501B parameters. Not a laptop model.
- Apache 2.0 allows commercial use, but training data won't be released ("proprietary licensed datasets").
What to do
- Don't plan a migration, plan an evaluation. Wait for the weights and technical report.
- Test on your own use case (your code, your agents) before trusting benchmarks.
- Estimate hosting cost: 501B parameters need multiple high-end GPUs even quantized. Inference providers may be cheaper.
- Read the final license when it ships.
Sources
- Reflection blog: https://reflection.ai/blog/introducing-beam
- SiliconANGLE (5/10/2026): https://siliconangle.com/2026/10/05/reflection-ai-debuts-open-source-beam-model-with-501b-parameters/
- 9to6ai: https://9to6ai.com/reflection-beam-501b-open-weight-model/
- The Terminal: https://theterminal.space/ai/reflection-beam-501b-weights-due-october
Want to evaluate open models for your company? KGA IT offers a free consultation: kga-it.com.