Reflection Beam 501B-A23B: what hardware it takes to self-host the new Apache 2.0 coding model

Reflection, a new US lab, has released its first model, Beam, as open weights on Hugging Face. The model card describes a sparse mixture-of-experts with 501.3B total and 23.4B active parameters, a 1M-token context, text in and text out, and an Apache 2.0 license. It targets coding and agentic work. Like Muse Spark Open and Kimi K3, it is “open” at a size most people can’t run at home. So the useful question is what it takes to serve it yourself.

The spec sheet

Item	Value (model card / Hugging Face)
Parameters	501.3B total, 23.4B active
Layers	52: 3 dense + 49 MoE, 256 experts, 8 routed per token
Attention	Gated GQA; 43 sliding-window layers (512 tokens) + 9 full-context layers
Context	1M tokens
Checkpoints	BF16, FP8, NVFP4, plus a DFlash2 speculative-decoding drafter
Reasoning effort	low, medium (default), high, xhigh, max
Knowledge cutoff	June 30, 2026
License	Apache 2.0

On Reflection’s own benchmark table, as published in Hugging Face’s launch post, the BF16 model scores 81.6% on SWE-bench Verified, 80.1% on Terminal Bench v2.1 and 90.6% on GPQA Diamond. The NVFP4 build stays within about a point on most rows. These are vendor-reported numbers.

How much GPU you need

Per Hugging Face, the BF16 weights are about 1 TB and NVFP4 cuts them to about 280 GB. NVFP4 only runs natively on Blackwell. Older GPUs such as the H200 should use the FP8 checkpoint.

• NVFP4 on Blackwell: two B200s (360 GB) hold the weights but leave little room for KV cache. Hugging Face recommends planning for four B200s. Two B300s, a full 8×B200 node or 8× RTX Pro 6000 also work.
• FP8 on Hopper: SGLang’s recipe uses 8× H200.
• BF16: the vLLM example runs tensor parallelism across 16 GPUs (two 8× H200 nodes). With Transformers expert parallelism, 16 H200s hold 78 GB of weights each. Eight would need 139 GB each, which doesn’t fit on a 141 GB H200.

The 1M context is cheaper than it sounds. Only the 9 full-attention layers keep KV for the whole sequence. vLLM’s Day-0 post puts that at 18 KiB per token in FP8, or about 18 GiB of FP8 KV for a full 1M-token context, before sharding.

Official launch commands

vLLM ships a dedicated image, vllm/vllm-openai:reflection-beam. Its example serves NVFP4 on four GPUs with -tp 4, --reasoning-parser beam, --tool-call-parser beam --enable-auto-tool-choice, --kv-cache-dtype fp8 and the DFlash2 drafter at 7 speculative tokens. The plain BF16 version from Hugging Face’s post is:

vllm serve reflection/Beam-1.0-501B-A23B --tensor-parallel-size 16 --served-model-name beam

For SGLang on Blackwell, the recipe launches reflection/Beam-1.0-501B-A23B-NVFP4 with --tp 4 --attention-backend trtllm_mha. On Hopper, it uses the FP8 checkpoint with --tp 8 --attention-backend fa3. Both use --context-length 1048576, DFlash speculative decoding and --tool-call-parser beam --reasoning-parser beam. Transformers support needs transformers>=5.20.1.

On speed, vLLM says coding agents optimized its implementation, cutting TP8 batch-1 decode from 8.95 ms to 3.30 ms per step. With the DFlash2 drafter, a single request reached 708 tokens/s on TP8, at an acceptance length of about 3.7, on a low-entropy code-heavy benchmark. vLLM notes that open-ended chat will see lower acceptance.

No Blackwell? Try it hosted first

Beam is on Hugging Face Inference Providers through Baseten, Fireworks and Together. Point an OpenAI client at https://router.huggingface.co/v1 with model reflection/Beam-1.0-501B-A23B:auto. There are also one-click Inference Endpoints recipes for FP8 and NVFP4. That’s the cheapest way to see whether Beam beats your current open model before you rent a node. Our Kimi K3 local guide walks through the same self-host-or-rent decision for an even larger open model.