An Air Gap Is a Pipeline, Not a Wall
An air gap is a one-way artifact pipeline, not just a blocked network. This explainer follows six stages, from mirroring dependencies outside to operating inside with no egress.
Loading post…
Hugging Faceben burtenshaw, merve, Pedro CuencaAnnouncement
Welcome Beam: a 501B open model you can own, customize, and scale
Summary
Reflection released Beam, a 501B-parameter mixture-of-experts model with 23B active parameters per token, under the Apache 2.0 license, targeting coding and agentic use cases. The model supports a 1M-token context window using hybrid attention with 43 sliding-window and 9 global layers, and a residual stream that scales sublayer outputs by 1/sqrt(k) to maintain constant magnitude. Its MoE router uses a group-limited scheme with 256 experts, 8 routed per token, and the BF16 checkpoint occupies about 1 TB. Benchmarks show strong performance: 81.6% on SWE-bench Verified, 97.7% on AIME 2026, and 90.6% on GPQA Diamond, with minimal degradation in FP8 and NVFP4 quantizations. The model ships with a Rust-based harness, day-0 support in transformers, vLLM, and SGLang, and includes SFT and RL recipes, with a $10,000 Hub credits competition for community projects.
Context
Beam is the first open model from Reflection, a new US AI lab, released on Hugging Face. It is a 501B MoE model designed for long-context reasoning and agentic tasks, with a focus on coding. The model is fully open under Apache 2.0, allowing customization and self-hosting.
Approach / What changed
The model uses a sparse mixture-of-experts architecture with 256 experts, 8 active per token, and hybrid attention with 43 sliding-window and 9 global layers. It includes a Rust-based harness for optimal performance and supports multiple quantization formats (BF16, FP8, NVFP4). The post provides recipes for inference, supervised fine-tuning, and reinforcement learning using transformers, vLLM, and SGLang.
Takeaways
Beam is a 501B MoE model with 23B active parameters per token, offering the capacity of a large model with the speed of a smaller one.
The model achieves strong results on coding and reasoning benchmarks, including 81.6% on SWE-bench Verified and 97.7% on AIME 2026.
Beam is fully open under Apache 2.0, with day-0 support in major inference libraries and a Rust-based harness for self-hosting.
FAQ
What hardware is needed to run Beam?
The NVFP4 checkpoint requires about 280 GB and runs natively on Blackwell GPUs; FP8 checkpoints are suitable for Hopper and Ampere GPUs. Two B200 GPUs (360 GB) can hold the weights.
How does Beam handle long contexts?
Beam uses hybrid attention: 43 sliding-window layers cache only 512 tokens each, while 9 global layers cache the full sequence, reducing KV cache growth to 36 KiB per token.
What quantization formats are available?
Beam checkpoints are available in BF16, FP8, and NVFP4 formats. NVFP4 stores most parameters in 4-bit precision for Blackwell GPUs, while FP8 is for older GPUs.
Source and freshness
- Source
- huggingface.co
- Published
- Added to Yomu
Summaries are generated from the original by a language model. Check the original before relying on it.
See the idea
Visual explainers
Keep reading
Related reading
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
The best engineering reads, once a week
One useful email. Unsubscribe anytime.