Reflection AI 開源 501B 參數 Beam 權重及 Mirror CLI 訓練工具
A Lab Just Open-Sourced Its Training Harness (Plus 501B Weights)
Reflection AI 本週開源 Beam 的 501B 參數權重及實際訓練工具 Mirror CLI，權重採用 Apache 2.0 授權。Beam 是稀疏混合專家模型，總參數為 501B、每個 token 啟用 23B，提供 BF16、FP8 和 NVFP4 三種權重格式。目前沒有獨立基準測試，文中的效能説法來自廠商資料。
Mirror CLI 隨 Beam 一同開源，讓讀者看到權重以外的訓練流程，並評估其對微調及預訓練研究的參考價值。
Reflection AI dropped Beam this week — 501 billion parameters of open weights under Apache 2.0. That's the headline. But here's what actually caught my eye: they also open-sourced Mirror CLI, their actual training harness. Meta never did that with Llama.
I haven't run Beam myself — no independent benchmarks exist yet, so everything below comes from Reflection's release materials and launch coverage. Treat the performance claims accordingly.
The model
Beam is a sparse mixture-of-experts: 501B total parameters, 23B active per token. It's aimed at coding, reasoning, and agentic workloads — the same territory as this week's Muse Spark Open (816B, also open-weight, very different license).
Three weight formats ship: BF16, FP8, and NVFP4. You get the tokenizer and a reference implementation too.
The license is the flex
Apache 2.0. Full stop. No attribution strings, no revenue thresholds, no "notify us before offering it as a service." If you want to self-host behind a firewall or fine-tune without a lawyer reading the license first, this is the one.
Compare that to Muse Spark Open's custom license (attribution required, $20M+ revenue companies need a commercial agreement). Same week, two 500B+ open models, opposite philosophies.
Mirror CLI: the part nobody's talking about
This is the genuinely unusual bit. Reflection open-sourced the actual training harness they used — Mirror CLI. Frontier labs don't do this. Meta's Llama releases gave you weights and a prayer; the training code stayed locked up.
Why it matters: if you're doing serious fine-tuning or pretraining research, seeing how a frontier lab actually orchestrates training runs is worth more than the weights themselves. It's the difference between getting a cake and getting the recipe plus the kitchen.
DFlash2
Beam also ships with DFlash2, a speculative-decoding drafter aimed at agentic latency — the stop-and-go pattern where agents make many small sequential calls. Worth watching if you're building agents, though again: vendor claims, no independent numbers yet.
Should you care?
Yes, if you:
- Want a permissive-licensed 500B-class model for self-hosting or fine-tuning
- Do training/fine-tuning research and want to see a real lab's harness
- Are comparing open-weight options this week (Beam vs Muse Spark Open is the matchup)
Probably not yet, if you:
- Need proven benchmark numbers (none independent exist)
- Want a model you can run on modest hardware (501B is still datacenter territory)
- Need hosted API pricing (free public beta for now, pricing not set)
My take
The weights are good news. The training harness is the story. A frontier lab handing over its actual training code is unprecedented in the open-weight world, and it's the kind of thing that moves the whole field forward — not just Reflection's users.
What I'd actually do: if you're evaluating open models this quarter, put Beam on the shortlist specifically for the license. Apache 2.0 removes an entire category of legal friction.
What open-weight release actually changed how you build — the weights, or the tooling around them?
Sources: Reflection AI announcement (Oct 15-16, 2026); launch partner coverage. All capability claims are vendor-reported; no independent benchmarks exist yet.
來源：DEV Community · dev.to