Reflection Beam Open Weights: What 501B Under Apache Offers

Reflection Beam is now something you can actually download. On October 16, US lab Reflection published the weights for Beam 1.0, its first open-weight model, under the permissive Apache 2.0 license. The model was introduced on October 5, but today's release is the one developers have been waiting for: full weights, several precision formats and day-zero support across the major open-source inference engines.

• Architecture: a sparse mixture-of-experts model with about 501B total parameters and 23B active per token, according to Reflection and Hugging Face.
• Context and reasoning: a 1M-token context window with five reasoning-effort levels, from low to max.
• Reflection's benchmarks (BF16): 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench 2.1, 90.6% on GPQA Diamond and 97.7% on AIME 2026.
• Try it now: Ollama Cloud runs it with one command, free for a limited time, and Reflection's own API is in public beta.

What Is Reflection Beam?

Beam is a large reasoning model built for coding and agentic work. The Hugging Face launch post explains that each token routes through 8 of 256 experts, which is how a 501B-parameter model runs with only 23B parameters active. That gives it the speed profile of a much smaller model, though every expert still has to fit in memory.

The long context is handled efficiently too. Hugging Face reports that at the full 1M-token context the key-value cache is about 36GiB per sequence, far less than it would be if every layer used global attention. A DFlash speculative-decoding drafter ships alongside the main weights to speed up generation further.

How Does Beam Score on Coding Benchmarks?

Reflection published results for each precision format, and the quantized versions hold up well:

• SWE-bench Verified: 81.6% (BF16), 80.2% (FP8), 80.7% (NVFP4)
• Terminal-Bench 2.1: 80.1%, 78.9%, 78.2%
• GPQA Diamond: 90.6%, 90.3%, 89.3%
• AIME 2026: 97.7%, 98.0%, 97.5%

These figures come from Reflection's own testing, so independent evaluations are the next thing to watch. Even so, an Apache-licensed model posting numbers in this range is a big step for open-weight coding agents, and it joins a busy autumn that also brought Mistral Large 4's 1T open-weight release.

What Hardware Do You Need to Run Beam?

This is data-center territory for self-hosting. Hugging Face lists about 1TB for the BF16 weights and about 280GB for NVFP4, with four B200 GPUs recommended for long contexts and concurrent requests. FP8 targets older Hopper and Ampere GPUs, and SGLang examples show FP8 on eight H200s.

For everyone else, hosted options are plentiful. Reflection lists AWS, Google Cloud, Microsoft Azure, Oracle, CoreWeave, Databricks, Cloudflare, Together AI, Fireworks AI, OpenRouter, LM Studio and NVIDIA NIMs at launch, with on-premises deployments through Dell and HPE. Ollama users can simply run the cloud version from a terminal.

Mirror CLI: An Open Agent Harness

Reflection also open-sourced Mirror, a terminal agent harness with a Rust core. It lets you approve each tool call, resume conversations, run parallel agents in separate worktrees and plug in skills and MCP connections. Releasing the harness alongside the weights means developers get a complete, inspectable coding-agent stack rather than a model alone.

Why Apache 2.0 Matters

Apache 2.0 is about as permissive as licenses get: companies can fine-tune, ship and sell products built on Beam without attribution banners or revenue thresholds. Together with day-zero support in Transformers, vLLM, SGLang, Unsloth and Axolotl, that makes Beam one of the easiest frontier-scale models to build on today. Keep up with the rest of the open-model race in our AI coverage.

Sources: Reflection: Beam is available — October 16, 2026; Hugging Face: Reflection Beam — October 16, 2026; Ollama: Reflection Beam — October 16, 2026; VentureBeat — October 16, 2026.