AI NewsModels & agentsAnnouncement
Reflection released Beam's weights on Hugging Face under Apache 2.0
Reflection published the full weights for Beam, a 501B-parameter Mixture-of-Experts model with 23B active per token and a 1M token context, under Apache 2.0, with day-zero support in transformers, SGLang and vLLM.
Image: Hugging Face
Why it mattersA team costing an open-weight model for coding or long-running agent work can now run Beam on its own hardware, fine-tune it, and ship a modified version without a licence call.
A team picking an open-weight model for coding or agent work has had to choose between models whose weights are either closed, licence-restricted, or trained in China. Reflection, a Brooklyn lab, just added a frontier-size Western option that teams can download and modify.
Reflection and Hugging Face published Beam's weights on 16 October 2026, eleven days after the "Introducing Beam" post that announced the model. Beam is a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active per token, a 1 million token context window, and an Apache 2.0 licence. The Hugging Face post calls it the first model from Reflection, built for software engineering, reasoning and long-horizon agent work.
What the release includes
Hugging Face published three checkpoint formats on the day: BF16, FP8, and NVFP4. The BF16 checkpoint is about one terabyte in safetensors; the NVFP4 checkpoint stores most weights in four-bit precision and shrinks that to around 280 gigabytes for NVIDIA Blackwell GPUs. Hugging Face says the four-bit numbers in its own table stay within 1 to 2 percentage points of the BF16 scores on the benchmarks it ran.
Day-zero inference support ships for transformers, SGLang and vLLM, with recipes for supervised fine-tuning in TRL and reinforcement-learning fine-tuning in Reflection's own Miles library. Hugging Face recommends four B200 GPUs for comfortable long-context serving on NVFP4, with the two-B200 and eight-B200 configurations as the lower and upper ends.
The numbers in the Hugging Face table
Hugging Face lists Reflection's own benchmark numbers in the BF16, FP8 and NVFP4 columns. On the BF16 checkpoint, Reflection reports 81.6 percent on SWE-Bench Verified, 76.5 percent on SWE-Bench Pro v2-Hard, 80.1 percent on Terminal-Bench 2.1, 97.7 percent on AIME 2026 and 90.6 percent on GPQA Diamond. These are the vendor's own figures and have not been checked by anyone outside Reflection.
The free trial on Ollama
Ollama is a day-zero partner and runs Beam on Ollama Cloud for free "for a limited time only". Ollama says the endpoint runs with zero data retention, works from the Ollama CLI and from Claude Code, Codex and OpenCode through the CLI's launcher. Ollama does not say when the free window closes or what the paid price will be, so a team planning past the pilot has to assume neither is set yet.
What else is in the kit
Hugging Face describes five reasoning levels in the tokenizer, from "low" through "medium" (the default) to "xhigh" and "max", and a response template that keeps a model's reasoning in context across turns through a preserve_thinking flag. Beam also ships with a dedicated open-source Rust-based harness that Hugging Face says delivers the model's best performance, and a speculative-decoding drafter model called DFlash2 for faster generation. Reflection is giving out $10,000 in Hub credits across the next few weeks for projects that build on Beam, which Hugging Face lists on the model's own page.
A team deciding whether to run Beam in-house gets more detail to cost it than most open-weight releases ship with: a per-sequence KV cache figure, a hardware table that names B200 and B300 options, and a note that only nine of Beam's 52 layers hold the whole context while the other 43 cap at a 512-token sliding window. The practical gate is still memory. The weights are out, the licence is Apache 2.0, and whether the vendor's benchmark numbers survive outside Reflection is the next thing a buyer finds out.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.
