Reflection unveils Beam, a 501B-parameter open-weight MoE model aiming to rival China's open models
SumitGup · x · 2026-10-06
Reflection has announced Beam, its first open-weight model: a sparse MoE with 501B total and 23B active parameters, built for coding, reasoning, and agentic workloads — pitched as the American answer to China's open-weight lead (GLM, Kimi, DeepSeek).
Key facts:
- Pretrained on 23.8T curated tokens (web + proprietary licensed data);
- High-compute RL run used 10.5K NVIDIA GB300 GPUs over 4 weeks, generating 100M+ rollouts;
- Claims competitiveness with GLM 5.2 and near-parity with Qwen 3.8-Max on coding/agentic tasks; its edge over Kimi K3 is inference efficiency rather than raw capability;
- Final red-teaming underway; weights, tech report, and model card due later this month, with early access open now.
Related event: Reflection AI Debuts Beam, a 501B-Parameter Open-Weight Model(20 posts)→
More from Models
- Dev argues Persimmon is the most underrated Western lab release: realistic user simulators may be quietly powering RL environments — willcb · 2026-10-06
- Developer hits Claude's 5-hour usage limit twice in one day — kevinkern · 2026-10-06
- beam called second-best Western model slated to go open-source — willcb · 2026-10-06
- Subscription math: Anthropic plans offer over 5X OpenAI's value vs equivalent API pricing — soumitrashukla9 · 2026-10-06
- SemiAnalysis: Opus 5.5 gives 5x the subscription value of GPT-6.1 Sol as OpenAI halves $200 plan limits — kimmonismus · 2026-10-06
- Bittensor guard model gains 8 F1 points in 4 weeks to near-SOTA via miner attacks — bittingthembits · 2026-10-06