Reflection AI Beam: 501B open-weight model for coding and agents
Reflection AI's first open-weight model: a 501B-parameter text-only MoE with 23B active parameters, 1M-token context, trained on 23.8T tokens with heavy RL over 1.3B sandboxes.
Why it matters
Claims parity with Chinese open models (GLM-5.2) at 3-4x less inference compute — the first serious Western open-weight answer to DeepSeek and Qwen.
What you could build with it
An indie developer or small consultancy could self-host Beam on rented GPUs to run a coding-agent service for enterprise clients, cutting inference bills to a fraction of closed-API costs while keeping client code entirely on-prem.
Does it hold up?
Too early to judge — weights, technical report and inference code ship later in October; no independent benchmarks exist yet, only Reflection's own numbers.
Built with Reflection AI Beam
- Ollama announces support for upcoming Reflection Beam launchlinkedin · Ollama publicly welcomed Beam and said it will support the upcoming launch of US open models, signaling day-one local runnability.
- Wccftech: 10,500 GB300 GPUs and 46.4M sandboxes/day to train Beamarticle · Breakdown of Beam's RL training: 1.3B sandboxes over 4 weeks, 1M coding/agentic/STEM environments, 100M rollouts, controllable length-penalty RL.
- Beam hits #2 on Hacker News front page (321 points)hn · HN front-page discussion of the launch, noting weights, technical report and Apache 2.0 license ship later in October.
Learn more
First spotted on article: source.
More AI models
ElevenLabs Eleven v4 + v4 Turbo: new emotive TTS architecture with a 100ms real-time variantElevenLabs released Eleven v4 and Eleven v4 Turbo on September 28: a new text-to-speech architecture with inline…models · JEV 0.72VoiceStudio: open-source, fully-local ElevenLabs alternative with voice cloning, dubbing and dictation in 646 languagesVoiceStudio is an open-source, fully-local voice AI studio: voice cloning from a 3-second sample, voice design, video…models · JEV 0.71Mistral Forge: enterprise platform for training and continuously improving proprietary modelsMistral AI launched Forge (Sep 28) — a full-lifecycle model training platform (pre-training, SFT, DPO/ODPO, RL…models · JEV 0.69DeepSeek V4-Flash official API: public beta with upgraded agent capabilities, Responses API and Codex supportDeepSeek's official V4-Flash is now a public beta API with massively upgraded agent capabilities - benchmark scores…models · JEV 0.69kev: open, trainable Jev-like family of small decision models on Qwen3.5/3.8A family of small decision models (0.8B to 27B) built on Qwen3.5/3.8 that reproduces Jev's prefill-only typed-decision…models · JEV 0.68GPT-Synopsys: OpenAI and Synopsys multi-year deal to build an AI model that operates EDA toolsOpenAI and Synopsys signed a multi-year agreement on Sep 30 to jointly develop GPT-Synopsys, a specialized model that…models · JEV 0.66
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.