Reflection launches Beam, a 501B open-weights model for coding and agentic work

Beam is the first model from Reflection, a new US lab, and it arrived on Hugging Face with Apache 2.0 weights from day one. The spec sheet: a 501B-parameter mixture-of-experts model with 23B active per token, a 1M-token context window, text in and text out, checkpoints in BF16, FP8 and NVFP4, and a focus on coding and agentic workloads.

The architecture in short

• 52 decoder layers. The first three use a dense MLP; the rest use a MoE block with 256 experts, of which 8 run per token. That is why a 501B model computes each token with about 23B parameters.
• Hybrid attention: 43 layers with a 512-token sliding window and 9 global layers that see the full context. The KV cache grows about 36 KiB per token, roughly a sixth of the naive rate for a full 1M context.
• Reasoning effort in five levels, from "low" to "max", plus a preserve_thinking option that keeps the reasoning across turns.
• Tool calling with its own grammar, and a response template in the tokenizer that splits reasoning, answer and tool calls without regex.

The benchmarks, as reported by Reflection

All the numbers below come from the Reflection team, so they are vendor-reported:

• SWEBench Verified: 81.6% (BF16).
• SWE Bench Pro v2-Hard: 76.5%; v1: 66.1%.
• Terminal Bench v2.1: 80.1%.
• AIME 2026: 97.7%; GPQA Diamond: 90.6%; LiveCodeBench: 88.3%.

Quality drops very little across precisions: going from BF16 to NVFP4, SWEBench Verified moves from 81.6% to 80.7%, and SWE Bench Pro v2-Hard actually improves, from 76.5% to 78.6%.

The hardware math

This is the part that decides whether the model is for you. The NVFP4 checkpoint compresses the 501B parameters from about 1 TB in BF16 to about 280 GB, but NVFP4 is native to Blackwell. The planning numbers from the post:

• 4x B200 (720 GB) for NVFP4, with room left for long contexts.
• 4x B300 for the SGLang recipe.
• 8x H200 with the FP8 checkpoint, for non-Blackwell hardware.
• 8x RTX Pro 6000 is listed as an option.
• No hardware of your own: serverless access through Hugging Face Inference Providers (Baseten, Fireworks, Together), or one-click Inference Endpoints.

Day-0 support covers transformers, vLLM and SGLang, all exposing an OpenAI-compatible server, plus post-training recipes: SFT with TRL, full-parameter GRPO reinforcement learning with Miles, and LoRA. Reflection is also putting $10,000 in Hub credits behind the best public project built with Beam.

What to take from it

Beam is an option aimed at teams that already run hardware and want a frontier-scale model they can own and fine-tune under Apache 2.0, for coding and agent work. If that is your situation, the numbers to verify before committing are the SWEBench and Terminal Bench results, and the NVFP4 memory math on your specific GPUs. If you are not running hardware today, the practical door is the serverless providers, where you can test the model before the hardware conversation starts.