Reflection Beam: 501B open weights, Apache 2.0, 80.9% SWE-bench
Nvidia-backed Reflection ships its first open model frontier-class on coding and free to try today; plus FLUX 3 Image Klein brings 4K image gen to an 8GB GPU.
Copy markdown
A frontier-class open model, now under Apache 2.0
Reflection's Beam is a 501B-parameter sparse MoE with 23B active and a 1M-token context, and the weights shipped today under a fully permissive Apache 2.0 license — no revenue cap, no gated terms. It's one of the strongest open models you can legally ship inside a commercial product without asking anyone's permission.
80.9% on SWE-bench Verified, 80.1% on Terminal-Bench
Reflection's own evals put Beam at 80.9% on SWE-bench Verified and 80.1% on Terminal-Bench v2.1 — agentic-coding scores that were closed-model-only a few months ago. An open model you can self-host at that level means your coding agent's brain no longer has to live behind someone else's API.
Pull the weights, or start free on the API
Weights ship in BF16, FP8 and NVFP4, with day-zero support in vLLM, SGLang, Transformers, Unsloth and Axolotl, so you can serve or fine-tune immediately. No GPU handy? There's a free public beta at platform.reflection.ai, and Baseten lists it at $1.30 per million input tokens and $0.13 per million output.
Mirror CLI ships alongside it: an open agent harness
Reflection also open-sourced Mirror CLI, a terminal coding-agent harness built around Beam that you run locally. It's a ready-made starting point if you'd rather build your own agent on open weights than wrap a closed CLI.
FLUX 3 Image Klein: 4K open image gen on an 8GB card
Black Forest Labs' new 5.6B open-weight image model does native 4K, pixel-perfect edits and bounding-box placement, with an INT8 build that runs on 8GB of VRAM (about 9s per 1MP image on an RTX 4070). ComfyUI support landed day-one and the API runs $0.009 per 1K images; its Kommunity license is free for commercial use under $5M in revenue.