Beam benchmarks: SWE-bench 81.6%, NVFP4 quantization barely hurts, fits 4×B200
ben_burtenshaw · x · 2026-10-16
The Hugging Face team published a hands-on deep dive on Beam, Reflection's first open model (501B MoE, 23B active, 1M context, Apache 2.0).
Specs: 52 layers, 256 experts with 8 routed per token; BF16/FP8/NVFP4 checkpoints; DFlash2 speculative decoding drafter; a dedicated Rust harness; preservethinking for cross-turn reasoning, tool calling with a dedicated grammar, and five reasoning levels (low to max). Day-0 support in transformers, vLLM, and SGLang, plus full-parameter RL via Miles.
Agent benchmarks:
- SWE-bench Verified: 81.6%
- Terminal Bench 2.1: 80.1%
- MCP Atlas: 80.8%
- NVFP4 quantization barely moves these scores — the 4-bit version fits on 4×B200.
The post also ships SFT and RL post-training recipes and end-to-end demos of running and customizing the model.
More from coding & agent
- vLLM ships Day-0 RL support for Beam in SkyRL, +62% LoRA decode throughput — vllm_project · 2026-10-16
- A 59-agent 'tui-review-and-fix' workflow completed in 1h18m, screenshot shows — max__drake · 2026-10-16
- Turning a Codex deep-dive article into a playable visual novel with every line sourced from transcripts — danshipper · 2026-10-16
- AI coding needs a new UI: separate chain of thought from current state and asks — robleclerc · 2026-10-16
- 'Best-of-n with resets' remains a hard-to-beat baseline for multi-agent workflows — mariofilhoml · 2026-10-16
- Troll tries to bait AI sales agent, gets a serious 3-page report — and a lesson in gating tool calls — Icy_Comfort_6220 · 2026-10-16