Reflection Launches Beam, a 501B Open AI Model With 1M Token Context
Reflection launches Beam, an Apache 2.0 501B open MoE model with 1M-token context, coding focus, day-one tooling and steep GPU requirements for teams.
Summary
On October 16, 2026, new US lab Reflection released its first model, Beam, on Hugging Face under Apache 2.0. The text-only coding and agentic model has 501B parameters, activates 23B per token, and supports 1M-token context, five reasoning levels, persistent thinking and grammar-based tool calls. Its 52 layers comprise three dense layers and 49 MoE layers, each with 256 experts but eight routed per token; experts hold 98.4% of parameters. Hybrid attention combines 43 layers using 512-token windows with nine global layers, reducing the 1M-token BF16 KV cache from 208 GiB to 36.08 GiB.
The BF16 checkpoint occupies about 1 TB; NVFP4 cuts weights to roughly 280 GB for NVIDIA Blackwell, while FP8 targets Hopper and Ampere. Two 180 GB B200s can hold NVFP4 weights, but four are recommended for activations, long contexts and concurrent requests. Reflection’s results show FP8 scoring 98.0% on AIME 2026, NVFP4 reaching 78.6% on SWE Bench Pro v2-Hard, and BF16 reaching 81.6% on SWE Bench Verified. Higher reasoning passed all five one-run task evaluations, while lower reasoning had formatting and justification failures on two.
Beam ships with a Rust harness, DFlash2 speculative decoding, and immediate Transformers, vLLM and SGLang support; Baseten, Fireworks and Together provide hosted inference. TRL LoRA tuning reached about 19,000 tokens per second on eight B300s and improved requested-language reasoning from 3 of 16 prompts to 16 of 16, while full-parameter Miles GRPO requires up to 96 H200s. Reflection is offering $10,000 in Hub credits for standout Beam spaces, models or datasets.
Positives
- Apache 2.0 licensing lets teams own, modify and deploy Beam on private hardware, infrastructure or cloud services.
- Eight routed experts activate only 23B of 501B parameters per token, preserving large-model capacity with lower computation.
- Hybrid attention reduces the 1M-token BF16 KV cache from 208 GiB to 36.08 GiB per sequence.
- FP8 Beam scored 98.0% on AIME 2026, while NVFP4 reached 78.6% on SWE Bench Pro v2-Hard.
- LoRA tuning raised requested-language reasoning adherence from 3 of 16 prompts to 16 of 16 after one epoch.
- $10,000 in Hub credits will reward selected Beam applications, quantizations, datasets, benchmarks or reinforcement-learning environments.
Risks & concerns
- All 256 experts remain resident, making the BF16 checkpoint about 1 TB despite only 23B parameters running per token.
- Four B200 GPUs are recommended for NVFP4 serving with useful room for activations, long contexts and concurrent requests.
- NVFP4 requires NVIDIA Blackwell hardware, forcing H200 and older GPU users toward FP8 or hosted inference.
- Full-parameter BF16 GRPO can require 96 H200 GPUs, including 64 for training and 32 for rollouts.
- Lower reasoning failed formatting on one long-context evaluation and omitted justification on a math task, despite otherwise correct content.
- Reflection supplied the benchmark results, while the five additional task evaluations were single-run vibe tests rather than formal benchmarks.