Reflection Beam: 501B Open Weights Under Apache 2.0
Reflection released Beam, a 501B MoE with 23B active and a 1M-token context, under Apache 2.0 on Oct 16. Per-precision benchmarks, GPU requirements, API pricing, and how its license differs from Meta's.
On October 16, a 501B frontier model landed under Apache 2.0
On October 16, 2026, Hugging Face published "Welcome Beam," and Reflection AI's weights became actually downloadable — eleven days after the October 5 preview. That makes Beam one of the rare cases where promised open weights arrived when promised.
The spec: a sparse MoE with 501B total parameters and 23B active per token, a 1M-token context window, text in and text out, under the Apache 2.0 license. It is aimed squarely at coding and agentic workloads. Baseten announced day-0 support the same day, AWS published a SageMaker JumpStart path, and 40+ partners were lined up at launch. What follows is primary-source numbers only.
What is actually inside
Three design choices matter, and Hugging Face documents all of them with numbers.
Mixture of experts. The first 3 layers are dense. The other 49 each hold 256 experts, with 8 routed per token using DeepSeek-V3's group-limited router — sigmoid over all 256, split into 8 groups of 32, keep the best 4 groups, then take the top 8 experts inside them. No shared expert. The experts hold 98.4% of the parameters.The practical consequence: all 256 experts stay resident in memory, so you need the memory of a 501B model but compute like a 23B one. Speed of a 23B, capacity of a 501B. Memory becomes the bill.
Hybrid attention. Of 52 layers, 43 use sliding-window attention over the previous 512 tokens, while 9 (every sixth plus the last) attend to the full context. Global layers use no positional encoding; instead each query is scaled as visible context grows,(1 + (pos + 1) / 2048) ** 0.1. It acts as a temperature so the softmax keeps precision instead of flattening out at length.
The payoff shows up in the KV cache. At 1M tokens: 208 GiB if all 52 layers were global, versus 36.08 GiB for Beam. At 128K: 26 GiB versus 4.6 GiB. Roughly one-sixth. Growth per token drops from 208 KiB to 36 KiB.
A constant-scale residual stream. A standard transformer adds every sublayer output with weight 1, so the stream grows with depth. Beam adds the output at step k with weight 1/√k and keeps the existing stream at √((k-1)/k). The squares sum to one, so after all 104 additions the input embedding and every sublayer output carry the same weight in the final hidden state.Benchmarks, published per precision format
Reflection's numbers are published in three columns — BF16, FP8, NVFP4 — which is unusual and useful, because it shows what quantization costs.
- SWE-bench Verified: 81.6% / 80.2% / 80.7%
- Terminal-Bench 2.1: 80.1% / 78.9% / 78.2%
- SWE-Bench Pro v1: 66.1% / 65.9% / 65.6%
- DeepSWE v1.1: 43.3% / 43.0% / 42.7%
- GPQA Diamond: 90.6% / 90.3% / 89.3%
- AIME 2026: 97.7% / 98.0% / 97.5%
- LiveCodeBench: 88.3% / 88.1% / 87.4%
- HLE (no tools / with tools): 36.0% / 51.1%
NVFP4 stays within roughly two points of BF16 almost everywhere, which is why Hugging Face calls it "the most practical way to serve the model."
Two caveats. First, every score is vendor-reported; nothing has been independently reproduced. Second, the numbers drifted between the October 5 preview and the release (DeepSWE v1.1 went 44.4 to 43.3; SWE-bench Verified went 80.9 to 81.6). Vendor tables move. Rerun your own tasks.
Also worth stating plainly: in Reflection's own comparison, Beam trails Chinese open models like Kimi K3 and GLM-5.3 on several coding benchmarks. The pitch is not "top of every chart." It is efficiency per token plus a US-origin Apache 2.0 license. Reflection claims GLM-5.2-level scores at 3-4x less inference compute, and reports hitting ~80% on Terminal-Bench 2.1 in roughly 25,000 tokens where GLM-5.2 needs about 80,000 for a similar score.
What it takes to run
This is the most practical section, because "memory is the bill" translates directly into requirements.
- BF16: about 1 TB — 16 H200s across two nodes
- FP8: roughly half — 8 H200s
- NVFP4: about 280 GB — 4 B300s. Two B200s (360 GB) hold the weights with almost no room for KV cache, so plan on four.
NVFP4 runs natively only on Blackwell; FP8 targets Hopper and Ampere. If you don't have that hardware, serverless is the simpler route.
Serving is a single command:
vllm serve reflection/Beam-1.0-501B-A23B --tensor-parallel-size 16 --served-model-name beam
transformers 5.20.1 or newer is required. Day-0 support covers transformers, vLLM, and SGLang, plus Unsloth, Axolotl, NVIDIA Dynamo, and TokenSpeed. Reflection also open-sourced Mirror CLI, the Rust-based harness it trained the model with — isolated workspaces, background sessions, and Slack, Linear, and GitHub integrations. A DFlash2 speculative drafter ships alongside it.
Pricing: where self-hosting stops paying
Baseten's model library lists Beam Model API pricing at $1.30 per million input tokens, $4.30 cache input, and $0.13 output. That $0.13 output number is striking — shaped for agents that think for a long time and answer briefly.
Against that, self-hosting still requires roughly 280 GB of fast memory at NVFP4 before the KV cache. Without sustained high volume, self-hosting does not pay off.
Compared to Meta's Muse Spark Open: different license shape
A day earlier, on October 15, Meta released weights for Muse Spark Open — 816B total / 42B active, 1M-token context, natively multimodal. That one is not Apache 2.0. It ships under the "Muse Model License," with three conditions:
So developers under $20M, internal use, research, and embedding it as a feature of your own product are free. Selling a hosted service is not. That is a different kind of constraint from Apache 2.0: Beam asks nothing, Meta imposes attribution plus a revenue threshold.
On performance, Meta's own table puts Muse Spark Open at 61.8 on JobBench (Inkling 31.1, Nemotron 3 Ultra 27.2), where Beam isn't listed. But MCP-Atlas is 89.2 for Meta versus 80.8 for Beam, while Terminal-Bench 2.1 favors Beam at 80.1 against Meta's 76.4. Different sponsors, different winners. Don't draw a conclusion from one table.
What this changes for you
Beam genuinely fits teams that run coding agents on their own infrastructure and cannot let data leave the network. Apache 2.0 means no contract negotiation and no attribution requirement if you commercialize.
It does not fit individuals on a laptop or single GPU. 280 GB at NVFP4; 1 TB at BF16. Wait for community quantizations. That tier is better served by running Muse Spark Open across two Mac Studio M5 Ultra machines (512 GB unified memory each, 1.2 TB/s bandwidth, ~1 TB combined), or by a smaller model entirely.
A short decision path:
- Want a one-week bake-off via API → Baseten and other inference providers. No hardware, from $1.30/$0.13
- Data must stay inside your network → 4 B200s or 8 H200s. Budget memory first
- Want to sell a hosted service → Beam is Apache 2.0; Meta requires a separate agreement past $20M. The license decides the business model
- Want to try it on a personal PC → neither works today. Wait for quantization
One more thing the benchmark tables hide: AI Weekly and VentureBeat report NVIDIA is in talks to acquire or deepen its stake in Reflection. Combined with NVFP4 being Blackwell-only, the picture is inference stack and silicon moving as one unit. Read only the score columns and you miss that.