Beam 1.0 (501B-A23B)
Reflection's first open model: 501B MoE with 23B active, 1M context, text only, Apache 2.0. Built for coding and agents, and claims about 3x fewer reasoning tokens than comparable open models. Hybrid sliding-window attention keeps the KV cache small (about 4.6 GiB per sequence at 128K in BF16).
Recommended quants
- Official NVFP4 (~294 GB)
reflection/Beam-1.0-501B-A23B-NVFP4The practical serving choice on Blackwell (4x B200, 2x B300, or 8x RTX Pro 6000 per the HF launch post). - Official FP8 (~503 GB)
reflection/Beam-1.0-501B-A23B-FP8For Hopper and older GPUs, e.g. 8x H200. - Unsloth UD-IQ1_S (103.1 GB) / UD-IQ2_XXS (133.9 GB) / UD-Q2_K_XL (186.9 GB)
unsloth/Beam-1.0-501B-A23B-GGUFLow-bit options for 128-256 GB machines; quality loss unmeasured here. - Unsloth UD-Q4_K_XL (306.5 GB) / Q8_0 (532.8 GB)
unsloth/Beam-1.0-501B-A23B-GGUFBig Mac Studio pairs or multi-GPU servers. - DFlash2 drafter
reflection/Beam-1.0-501B-A23B-DFlash2Speculative-decoding draft model; recommended in the official SGLang recipe.
Chat template Works
Modified Harmony format. The response template ships in the tokenizer, so transformers' parse_response splits reasoning, answer and tool calls for you. reasoning_effort: low, medium (default), high, xhigh, max. preserve_thinking keeps reasoning across turns. Tool-call arguments use Beam's own XML-style grammar.
Fix: Needs transformers >= 5.20.1. On SGLang pass --tool-call-parser beam --reasoning-parser beam, or tool calls and reasoning come back as raw text. When parsing tool calls in transformers, pass tools to parse_response too so arguments get cast to their declared types.
Sampling settings
| Official (generation_config.json) | temp 0.7, top_p 0.99, repetition_penalty 1.0 |
|---|---|
| Reasoning effort | medium by default; raise to high for hard coding or long-context retrieval (HF's vibe tests saw low effort miss steps) |
Runtimes
| vLLM | Works | Day-0 support. BF16 needs 16 GPUs (2x 8x H200): vllm serve reflection/Beam-1.0-501B-A23B --tensor-parallel-size 16. Optional DFlash speculative config. |
|---|---|---|
| SGLang | Works | Day-0 support. NVFP4 on 4x B300 (trtllm_mha) or FP8 on 8x H200 (fa3), with DFLASH speculative decoding and beam parsers. |
| llama.cpp | Unverified | Unsloth GGUFs published on launch day; minimum llama.cpp build not checked. |
| Ollama | Unverified | |
| LM Studio | Unverified | |
| MLX | Unverified |
Tool calling in agent tools
| Claude Code (local endpoint) | Unverified | |
|---|---|---|
| OpenCode | Unverified | |
| Cline | Unverified | |
| Aider | Unverified |
Known issues
Sources checked
- https://huggingface.co/blog/reflection-beam
- https://huggingface.co/reflection/Beam-1.0-501B-A23B
- https://huggingface.co/unsloth/Beam-1.0-501B-A23B-GGUF
- https://github.com/huggingface/transformers/releases
Spotted something wrong or a new fix? Corrections are reviewed before publishing. (Submission form coming soon.)