Reflection Drops Beam: A 501B Open-Weight Model Built for Inference Efficiency
Beam packs 501B parameters but only uses 23B active. It matches larger models on coding while slashing inference costs by up to 4x.
Models · Source: Hacker News
What happened
Reflection just announced Beam. It is a 501 billion parameter open-weight model with a sparse mixture-of-experts architecture. Only 23 billion parameters are active during inference. The model focuses heavily on coding, reasoning, and agentic workflows. Reflection designed this to be a workhorse for enterprise applications.
The training scale is massive. Reflection pretrained Beam on 23.8 trillion diverse tokens from the web and proprietary datasets. Then they hit it with a massive reinforcement learning run using 10.5K NVIDIA GB300 GPUs over four weeks. This generated over 100 million rollouts across a million curated environments. They used 1.3 billion sandboxes to grade the model.
Beam matches models like GLM 5.2 on advanced reasoning but uses three to four times less inference compute. It approaches Qwen 3.8-Max on coding tasks. While frontier models like Kimi K3 still lead on raw capability, Beam wins on efficiency. The weights, model card, and technical reports drop later this month.
Key facts
- 501B — Total parameters in Beam
- 23B — Active parameters per token
- 23.8T — Tokens used for pretraining
- 10.5K — NVIDIA GB300 GPUs used for reinforcement learning
- 100M — Reinforcement learning rollouts generated over 4 weeks
- 256K — Maximum context length in tokens
- 3-4x — Less inference compute used compared to GLM-5.2
Why it matters
This changes the math for founders building agentic workflows. Agents require thousands of API calls and massive context windows to solve real problems. Running a dense 500 billion parameter model for multi-step reasoning burns cash fast and destroys unit economics. Beam gives you frontier-level coding capabilities at a fraction of the hardware cost because it only activates 23 billion parameters per token. You get more intelligence per token without the massive infrastructure bill.
The second-order effect is a shift in how open labs approach reinforcement learning at scale. Reflection proved you can scale asynchronous reinforcement learning without policy staleness wrecking the training run. They maintained stable learning even when interactions were generated a day earlier and 107 weight versions behind. We will see more open-source labs copying this infrastructure. They will invest heavily in synthetic data pipelines and massive reinforcement learning runs to squeeze maximum intelligence out of smaller active parameter counts.
For builders
Dial in your reasoning costs
Beam includes a controllable reasoning effort parameter. You can force shorter responses for cheap tasks or allow longer reasoning for complex code. Builders save money by matching compute exactly to the task difficulty. You stop paying for unnecessary tokens.
Run massive agentic loops locally
The 23 billion active parameter count means you can run this efficiently on standard enterprise hardware. You do not need to pay massive API fees to frontier labs for high-volume agentic tasks. Startups building autonomous coding agents win big here. Incumbents selling expensive API wrappers lose their margin advantage.
Wait for the actual weights
Reflection is only offering early access right now. The weights drop later this month alongside the developer artifacts. Do not rip out your current inference stack until the open-source community verifies these benchmarks. Wait for the actual release to test the context window limits.
My take
I love seeing open-weight models attack inference costs instead of just chasing raw benchmark scores. Startups die paying for compute, not from a lack of intelligence. Beam proves that massive reinforcement learning runs on sparse architectures are the future of profitable AI products.
Original reporting: Hacker News. This is my rewrite and opinion.