Two agent models went open-weight in 24 hours: Reflection's Beam under Apache 2.0, StepFun's Step 5 Preview under its own license

Reflection AI promised Beam's weights for later this month when it announced the model on October 5. They landed on October 16 at about 13:00 UTC, under Apache 2.0, in BF16, FP8 and NVFP4, with a speculative-decoding drafter alongside. The Hugging Face repo had 440 downloads and 95 likes within hours and Unsloth had a GGUF up in under an hour. Beam is the 501B mixture of experts with 23.4B active, 1M context, text only, gated GQA plus sliding window attention, a modified Harmony chat template and a June 30 knowledge cutoff. Reflection says it reaches comparable solution quality to leading open models with about three times fewer reasoning tokens, which has been the pitch since day one.

The model card numbers moved a little from the launch post and the quantized versions are the interesting part. DeepSWE v1.1 is 43.3 percent in BF16, 43.0 in FP8, 42.7 in NVFP4. Terminal-Bench 2.1 is 80.1, 78.9, 78.4. SWE-Bench Verified is 81.6, 80.2, 80.7. And on SWE-Bench Pro v2-Hard the 4-bit NVFP4 checkpoint scores 78.6 against 76.5 for full precision. Reflection is still well behind Kimi K3 on DeepSWE and says so, and Meta's Muse Spark Open table a day earlier had Beam at 78.7 on MCP-Atlas against Muse's 89.2.

What Reflection is actually launching is a distribution. More than 40 partners on day zero: AWS, Azure, Google Cloud, Oracle, Databricks, Cloudflare, CoreWeave, Baseten, Fireworks, Together, Modal, Nebius, Nvidia NIMs, Ollama, LM Studio, OpenCode, OpenRouter, Vercel and Weights and Biases, plus fine-tuning and RL paths through NeMo RL, TRL, SkyRL, Axolotl and Miles. Reflection calls this an open operating system for agents where the model is the kernel. It also open-sourced Mirror CLI, the harness Beam was trained with: a Python front end for prompts, skills and tooling, talking to a Rust agent core that drives the model, with worktree-isolated parallel agents, resumable sessions, MCP connections and Python extensions. Reflection says Mirror started as an internal tool so agents could keep working after engineers closed their laptops, and that the biggest gains were in code review. The ecosystem design is the tell: a US open-weights lab that lobbied Washington in May about Chinese open models being a national security problem is now selling sovereign deployment to the buyers who care about that.

StepFun did the same thing a day earlier with less noise. Step 5 Preview, announced September 20 with a promise of weights on October 15, went up on Hugging Face and ModelScope on schedule in BF16 and FP8, with an NVFP4 repo created the day before. It is a 600B mixture of experts with 27B active, 1M context, text plus image plus video input, aimed at software engineering, professional work and specifically finance. StepFun's own chart has it at 67.7 on DeepSWE v1.1 in the SWE-agent harness against 67.5 for Kimi K3 and 66.9 for GLM 5.3, 80.5 on ProgramBench, 66.4 on FrontierFinance, 29.5 on Agents' Last Exam, and 33.3 on Terminal-Bench v4 where GLM 5.3 leads at 41.9. The license is StepFun Model License 1.0, which the card describes as Apache-2.0-style with additional attribution and commercial-license conditions. That is three open-weight agent models in two days, Meta's gated, StepFun's conditional, Reflection's plain Apache, and the license column is now part of the scoreboard.

Links: https://reflection.ai/blog/beam-is-available , https://reflection.ai/beam-model-card , https://huggingface.co/reflection/Beam-1.0-501B-A23B , https://developers.reflection.ai/mirror-overview , https://huggingface.co/stepfun-ai/Step-5-Preview , https://www.stepfun.com/step-5-preview , https://x.com/StepFun_ai/status/2110717330657616092