What Is Reflection Beam? The 501B Open-Weight Model (2026)
The short answer
Reflection Beam 1.0 is a US-built, Apache 2.0 open-weight coding and agent model: 501.3B total parameters, 23.4B active, 1M context, text only. Reflection AI first announced it with early API access and released the weights, model card and safety report on Hugging Face in mid-October 2026 (OpenRouter lists the beam-1.0-501b-a23b-20261016 build). Its selling point is token efficiency: Reflection says Beam matches leading open models with about 3x fewer reasoning tokens, which lowers cost and latency per task more than its raw scores suggest.
Key facts
| Beam 1.0 (501B-A23B) | |
|---|---|
| Developer | Reflection AI |
| Licence | Apache 2.0 (commercial use, modification, redistribution) |
| Architecture | Sparse MoE, 52 layers (3 dense + 49 MoE), 256 experts, gated GQA + sliding-window attention |
| Parameters | 501.3B total · 23.4B active |
| Context | 1M tokens |
| Modalities | Text in, text out |
| Knowledge cutoff | June 30, 2026 |
| Weights | BF16, FP8, NVFP4 + speculative-decoding drafter |
| Training | 23.8T pretraining tokens; RL with 100M+ rollouts on 10.5K NVIDIA GB300 GPUs for four weeks |
| Harness | Mirror CLI (Reflection’s open coding agent), plus OpenCode and other harnesses |
Benchmarks (vendor-reported, BF16)
| Benchmark | Beam |
|---|---|
| SWE-bench Verified | 81.6% |
| SWE-bench Multilingual | 78.3% |
| Terminal Bench v2.1 | 80.1% |
| SWE-Bench Pro v1 | 66.1% |
| DeepSWE v1.1 | 43.3% |
| GPQA Diamond | 90.6% |
| AIME 2026 | 97.7% |
| HLE (with tools) | 51.1% |
| MCP Atlas | 80.8% |
| BrowseComp (context management) | 76.8% |
FP8 and NVFP4 scores are within about one point of BF16 on most rows, so the quantised builds are the practical default. In Reflection’s launch table, DeepSeek V4.1 Flash (74.2), Kimi K3 (68.0) and GLM-5.3 (61.0) score higher on DeepSWE v1.1; Reflection argues Beam reaches GLM-5.2-level reasoning with 3–4x less inference compute. Treat all of these as the vendor’s numbers until independent evaluations settle.
Where to run it and what it costs
| Option | Price (per million tokens, Oct 21, 2026) | Notes |
|---|---|---|
| Nebius via OpenRouter | $0.25 in / $1.00 out | FP8, 1M context, 131K max output |
| CoreWeave via OpenRouter | $0.30 / $1.00, $0.04 cached | FP4, implicit caching |
| Baseten, Together via OpenRouter | $0.40 / $1.20, $0.04 cached | FP4 / unspecified |
| Reflection API | Free-to-start public beta | platform.reflection.ai |
| Self-host | Your GPUs | ~1 TB BF16, ~500 GB FP8, ~250–300 GB NVFP4 for weights alone (our estimate from parameter count) |
Reflection also lists AWS, Google Cloud, Microsoft Foundry, Databricks, Cloudflare, Fireworks, Modal, NVIDIA NIM, Oracle, Vercel, Ollama and LM Studio among launch platforms, and Dell and HPE for on-prem and air-gapped installs. For context, Beam’s cheapest hosted rate is in the same band as DeepSeek V4.1 Flash off-peak ($0.15 / $0.60) and far below Claude Sonnet 5.5 ($2 / $10) — see current API prices.
Who should care
- Teams that need a permissive US open-weight model for coding agents on their own infrastructure: Apache 2.0 is simpler than custom model licences.
- Cost-sensitive agent workloads where tokens per task matter more than peak benchmark scores.
- Not for multimodal work — Beam is text only; for images or video look at Muse Spark Open or Nemotron 3.5 Super VL.
For the wider field of self-hostable models, see the best open-weight AI models to self-host; for the company, what is Reflection AI.
Last verified: October 21, 2026.