Reflection AI Open-Sources Beam: A 501B Model Under Apache 2.0 You Can Actually Ship

What Just Landed on Hugging Face

Two weeks ago, Reflection AI — the startup founded by DeepMind alums — was showing benchmark charts for a model nobody could download. On October 16, the weights landed: Beam, a 501-billion-parameter mixture-of-experts model built for coding, reasoning, and long-horizon agentic work, published to Hugging Face under Apache 2.0.

The specs are frontier-class. Beam is a sparse MoE: 501B total parameters, but only ~23B active per token, which is what keeps inference tractable. It has an effective context window of 1 million tokens, ships with a tokenizer, reference implementation, and an agent harness, and is available in BF16, FP8, and NVFP4 variants with a DFlash2 speculative decoder. Reflection says the weights fit on a single eight-GPU server in 8-bit or 4-bit quantization — not a data center, a rack. Alongside the model, the company also open-sourced Mirror CLI, the training harness that built it, which may end up mattering more to researchers than the weights themselves.

The headline benchmark claim: Beam scores roughly 80% on Terminal-Bench 2.1 while spending about 25,000 tokens per attempt — versus GLM-5.2 needing around 80,000 tokens for similar scores. Those are Reflection's own numbers from its own harness, so treat them as a claim to verify, not a settled fact. But the pattern matches what the training run was optimized for (more on that below).

The License Is the Real Story

Plenty of "open" models in 2026 are open the way a museum is open: you can look, you can visit, but there are rules on the wall. Community licenses come with user caps (Meta's Llama-era terms), naming clauses, usage restrictions, and downstream-training limits that make any legal team slow-walk adoption for weeks.

The timing is what makes Beam interesting. On Thursday, Meta shipped Muse Spark Open — 816B parameters, genuinely capable — under a community license with conditions attached. On Friday, a much smaller lab shipped a frontier-scale model under a license written in 2004 with none. As coverage noted this week, the open-weight contest is no longer a fight over parameter counts; it's a fight over what you're allowed to do with the file.

Apache 2.0 drops the user caps, attribution theater, and downstream limits entirely. For anyone evaluating open weights for production, the legal review just went from a month to an afternoon. That's a deliberately cheap competitive lever for a lab that can't outspend Meta on compute — and it puts pressure on every frontier lab still attaching homework to their "open" releases. DeepSeek's MIT-licensed checkpoints and now Beam's Apache 2.0 make community-license holdouts look increasingly defensive.

The Token-Efficiency Trick

The most quietly important detail in Beam's training story is how it was optimized. Reflection ran a four-week training schedule on 10,500 Nvidia GB300 GPUs with more than 100 million reinforcement-learning rollouts — and the reward function penalized unnecessary tokens.

That's a direct response to how agentic models actually fail in production. When a coding agent burns 80,000 tokens grinding through a task, you pay for every one of them, twice — in API cost and in wall-clock time. A model that reaches the same answer in a third of the tokens isn't a modest efficiency gain; it's a 3x cut to your per-task bill and faster feedback loops for every agent built on top. This is the same bet DeepSeek made with its Flash line: in 2026, token efficiency is a feature users can feel, not just a benchmark footnote.

It also changes the effective hardware story. Fewer tokens per task means an eight-GPU server running Beam serves more agent sessions per hour than raw parameter counts suggest. For teams self-hosting coding agents — whether via Claude Code, OpenAI Codex, or open harnesses — that math is the difference between a pilot and a product.

How to Actually Run Beam

Beam arrived with unusually deep day-0 ecosystem support — 40+ partners, which matters because a model nobody can serve is a research artifact, not a tool:

• Ollama is offering Beam free for a limited time on Ollama Cloud with US hosting and zero-data retention — ollama run beam, or launch it straight into Claude Code and Codex via ollama launch claude --model beam:cloud.
• vLLM shipped day-0 support for the weights, with day-0 RL support landing in SkyRL for anyone wanting to fine-tune agentic behavior.
• AWS published a deployment guide covering SageMaker JumpStart and HyperPod Inference, including capacity planning.
• Self-hosters get BF16, FP8, and NVFP4 checkpoints plus the DFlash2 drafter for faster inference on modest hardware.

If you've been waiting to experiment with a frontier-scale open model without negotiating a license or renting a cluster, this is the lowest barrier of entry so far this year.

Beam in the 2026 Open-Weight Race

Beam lands in the middle of the most crowded open-weight week of the year. How the major releases stack up:

The takeaway from the table isn't which row wins — it's that the rows now differ mainly on terms and hardware, not raw capability. NVIDIA's Bolt puts a 74B agent model on a desktop card; DeepSeek's Flash checkpoint costs almost nothing to run; Meta's Muse Spark Open is the biggest but comes with strings; Beam is the one your lawyer clears in an afternoon. The frontier-lab vs. startup gap, on both legality and practicality, just narrowed again.

What It Means for the Tools You Pick

• License terms are now a buying criterion. When two open models are close in quality, Apache 2.0 beats community license every time for anything commercial. Put the license next to the benchmark in your eval spreadsheet.
• Token efficiency compounds. For agentic workloads, cost-per-completed-task beats cost-per-token. Ask vendors for Terminal-Bench-style numbers with token spend, not just accuracy.
• Portability is real now. With Beam launchable into Claude Code, Codex, and OpenCode from Ollama in one command, the coding-agent layer is decoupling from the model layer. Tools like Cursor and GitHub Copilot increasingly compete on workflow while models become swappable.
• Expect independent reproductions before trusting the charts. Reflection's numbers are self-reported from its own harness. The interesting next two weeks are third-party evals — and whether any frontier lab responds by loosening its own license terms.

Beam won't dethrone the biggest closed models this quarter, and 501B parameters is still serious hardware for most teams. But as a signal of where open-weight AI is heading in 2026 — permissive by default, efficient by design, deployable everywhere from Ollama to SageMaker — it's one of the cleanest statements yet. The models that get used are the ones a legal team clears quickly and hardware you already own can hold.