Reflection unveils Beam, a 501B open-weight MoE trained with RL on 10.5K GB300 GPUs
thione · x · 2026-10-12
Reflection introduced Beam, its first open-weight model: a sparse MoE with 501B total / 23B active parameters built for coding, reasoning, and agentic workloads.
- Pretrained on 23.8T curated tokens, then scaled high-compute RL: over 100M rollouts on 10.5K NVIDIA GB300 GPUs across 4 weeks.
- Claims it advances the Western open-weight frontier, competitive with GLM 5.2 and approaching Qwen 3.8-Max on coding/agentic tasks; its edge over Kimi K3 is inference-time efficiency.
- Final red-teaming underway; early access open, with weights, tech report, and model card due later this month.
The same recap notes Mistral previewed Large 4, a trillion-parameter model focused on cyber/enterprise work, also with open weights planned for late October.
More from Models
- Leak: Meta's Fable-class Muse Wonder model rolls out at $4 in / $20 out per Mtok — thesaraharminta · 2026-10-13
- Baidu Qianfan personal token plan: 'no peak throttling' but 429s every minute — NeuroSaki987 · 2026-10-13
- GPT-6.1 Sol Matches Flagship Within 1 Point at 1/5 the Price; GPT-6.1 Astra Pulled Over Deception — DeepLearningAI · 2026-10-13
- AI memory tested: GPT-4o drops from 87% to 60.6% when fed the full 115k-token chat history — gaganghotra_ · 2026-10-13
- MoonMapBench: Opus 5.5 Off by 1,192 m Predicting Lunar Elevations — maksym_andr · 2026-10-13
- GPT-6.1 Sol Praised as Fast Workhorse, But Daybreak Mode Blocks Manual Model Picks — rez0__ · 2026-10-13