Reflection 開放 Beam 權重：每 token 啟用 23B，部署仍須容納 501B 參數

Reflection released the weights for Beam on October 16, 2026. It is the lab's first open-weight model: a sparse mixture-of-experts (MoE) model with 501B total parameters, about 23B active per token, a 1M-token context window, and an Apache 2.0 license.

Most posts stopped at the headline numbers. The part that matters if you plan to run it yourself is the gap between "23B active" and "501B on disk". Here is what was released, why people care, and what the coverage skipped.

What happened (facts)

The model. 501.3B total and 23.4B active parameters, 52 layers (3 dense + 49 MoE), 256 experts per MoE layer with 8 routed per token, gated GQA plus sliding-window attention, a 152k vocabulary, text in and text out, predominantly English. Knowledge cutoff: June 30, 2026 (model card).
The files. Instruction-tuned checkpoints in BF16, FP8 and NVFP4, plus a speculative-decoding drafter model.
The license. Apache 2.0. The model card says Reflection does not prohibit distillation, training other models on outputs, fine-tuning, quantization or merging.
Day-zero support. Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and NVIDIA Dynamo, plus hosted access through Reflection's API and partners including AWS, Google Cloud, Cloudflare, OpenRouter, Together AI and Vercel.
The harness. Reflection also open-sourced Mirror CLI, the coding-agent harness it trained Beam alongside: a Python frontend talking to a Rust agent core.

Why people care

Apache 2.0 with no field-of-use clauses is the cleanest license a frontier-sized open model can ship with. Compare that with "open-weight" licenses that add user caps or acceptable-use terms you have to read line by line. For a SaaS team, Apache 2.0 means you can fine-tune it, quantize it and serve it inside your product without a separate commercial agreement.

The other reason is cost per token. Only about 23B parameters run for each token, so once the model is loaded, generation compute is closer to a mid-size dense model than to a 500B one. Reflection also claims Beam reaches comparable quality to leading open models with about 3x fewer reasoning tokens.

What it could change (prediction, not fact)

If the token-efficiency claim holds up in independent tests, Beam becomes a realistic self-hosted backend for coding agents, where long reasoning traces are the main cost. That is a prediction. The technical report is listed as "coming", and the benchmark table on the model card is Reflection's own run (temperature 0.7, top_p 0.99).

What the coverage misses

1. "23B active" is a compute number, not a memory number. Every expert has to sit in memory, because any token can route to any of them. Rough weight-only math, our estimate rather than a published figure:

BF16 (2 bytes per parameter): about 1 TB
FP8 (1 byte): about 500 GB
NVFP4 (about 4.5 bits with scales): roughly 280 GB

That is before the KV cache, and a 1M-token context makes the KV cache the second big bill. Sliding-window attention on some layers helps, but plan for multi-GPU servers, not a single card. For most teams, a hosted endpoint is the practical way to try it first.

2. The quantized checkpoints are official. Reflection published scores for BF16, FP8 and NVFP4 side by side, and they sit within about a point of each other on most benchmarks (SWE-bench Verified: 81.6%, 80.2%, 80.7%). That is more useful than a community quant with no evals.

3. Benchmarks are self-reported for now. SWE-bench Verified 81.6%, Terminal Bench v2.1 80.1% and GPQA Diamond 90.6% are strong numbers, but they come from Reflection's own evaluation. Wait for third-party runs on your kind of task before switching anything in production.

4. The chat template is a modified Harmony format. If your serving stack hard-codes another template, tool calls are the first thing to break. Use the template shipped with the weights.

Before you self-host Beam

Try a hosted endpoint on your real prompts first.
Pick the precision you can afford in memory, then check its row in the model card's table.
Budget KV cache for the context length you actually need, not the 1M maximum.
Use the official chat template and test tool calling early.

Sources

Reflection, "Beam is now available" (October 16, 2026): https://reflection.ai/blog/beam-is-available
Model card: https://huggingface.co/reflection/Beam-1.0-501B-A23B
Hugging Face blog: https://huggingface.co/blog/reflection-beam