Reflection AI Launches Beam, Its First Open-Weight Model: 501B Total Params, 23B Active
On October 6, Reflection AI, an Nvidia-backed NYC startup founded by former DeepMind/Google researchers (CEO Misha Laskin), announced Beam, its first open-weight model: a text-only sparse MoE with 501B total and just 23B active parameters, positioned as a workhorse open model for coding, agents, and science. Full weights arrive this month under Apache 2.0, with documentation and a full stack of run/eval/finetune tooling. The team claims 3-4x token efficiency over GLM 5.2; the model was still in final red-teaming at announcement, so real-world performance awaits community benchmarks.
Confirmed
- Sparse MoE architecture, 501B total / 23B active parameters, trained end-to-end from scratch on 23.8T curated tokens over roughly a year; per @rohanpaulai, pretraining took 4 weeks on 6,144 GB300-class GPUs
- Per reposts (@Madisonkanna), efficiency comes from RL length penalties suppressing redundant tokens plus the sparse MoE design; per @brandondamos, the team spent a year building a production RL/OPD stack stable on 10,000 GB300s; @AccBalanced cites 10.5K GB300s producing over 100M rollouts
- Positioned as advancing the Western open-source frontier on coding and agentic tasks, widely framed as a response to Chinese open models (GLM, Kimi, DeepSeek, Qwen)
- Apache 2.0 weights due this month with docs and tooling, targeting enterprises, governments, and developers
- Engineer alexpolozov publicly thanked the pretraining team and leadership (Misha Laskin, Ioannis Antonoglou)
- Third-party observations (@gethackteam, @iaziaz) place Beam's public benchmark scores between Qwen 3.8 and GLM 5.3, competitive with GLM 5.2, though accounts differ on exact ranges
- @TansuYegen argues the story is not the 501B headline but that firms are learning to activate only a small slice of huge models per task; if open models keep getting smarter without runaway inference costs, closed-model moats erode
Unconfirmed
- The 3-4x efficiency claim is self-reported and lacks independent verification; @nmasc flags that specs and benchmark results await confirmation, and describes the founders as ex-DeepSeek/Google, inconsistent with the more common ex-DeepMind/Google account
- A reported $40M training cost (@Madisonkanna repost) and a $25B valuation mention (@WebAssemblyMan) are unverified; one repost (@AccBalanced) spells the CEO's name as Misha Likhov, inconsistent with Misha Laskin
Why it matters
Beam marks Reflection AI's public debut at 500B scale, directly benchmarking against GLM, Kimi, DeepSeek, and Qwen, and is widely read as the West's answer to China's open-source lead. Per alexpolozov (relayed by @bhutanisanyam1), the pretraining team was only 5-10 people when he joined last November, with no crawlers, GPU infrastructure, dedup pipelines, or internal evals — all built from scratch. Completing pretraining within a year underscores how fast the open-source frontier race is accelerating; the Apache 2.0 release this month makes real availability and measured performance worth watching.
2026-10-06 ~ 2026-10-07 · 38 related posts
- Episode 1: Nvidia-backed Reflection to release first open-weight model, taking on Chinese and US AI rivals(2026-10-04, 15 posts)
- Episode 2: Reflection AI Launches Beam, Its First Open-Weight Model: 501B Total Params, 23B Active(2026-10-06, 38 posts)
- Episode 3: Reflection's open-weight Beam debut met with benchmarks showing it trails Chinese open models(2026-10-06, 5 posts)
Primary sources
- Reflection unveils Beam: 501B-param open MoE model with 23B active, trained on 23.8T tokens — bigblueboo ·
- Reflection ships Beam, its first open-weight model, after a year of building pretraining from scratch — bhutanisanyam1 ·
- Reflection unveils Beam: 501B-param open model with 23B active, built for agentic efficiency — alexpolozov ·
- [source] Reflection unveils Beam: 501B-param open MoE model with 23B active, trained on 23.8T tokens — bigblueboo · 2026-10-06
- Reflection AI to open-source 500B agentic model Beam under Apache 2.0 this month — bigblueboo · 2026-10-06
- Reflection AI releases its long-awaited open model — nmasc_ · 2026-10-06
- [source] Reflection ships Beam, its first open-weight model, after a year of building pretraining from scratch — bhutanisanyam1 · 2026-10-06
- Beam ships first model: 501B-A23B text MoE openly released under Apache 2.0 — brandondamos · 2026-10-06
- Beam: 501B Open Model With 23B Active, Full Weights This Month, 3-4x More Efficient Than GLM 5.2 — TorturedPoet30 · 2026-10-06
- Reflection AI researcher pens farewell to Beam team, hailing it as a "GLM5.2-level Western open agentic coding powerhouse" — alexpolozov · 2026-10-06
- Reflection AI Open-Sources Beam: 501B-Param Sparse MoE With 23B Active, ~$20–40M RL Training — Madisonkanna · 2026-10-06
- Reflection's 501B Beam model touted as open, but weights missing and generalization test challenged — bigdata · 2026-10-06
- Beam launches: Mira's new frontier lab credits its first-ever pretraining team — alexpolozov · 2026-10-06
- Reflection AI's alexpolozov thanks the large pretraining team behind Beam — alexpolozov · 2026-10-06
- Reflection's 501B Beam: sparse activation is making open models a tougher economic rival — TansuYegen · 2026-10-06
- Reflection AI Claims Beam Is 3-4x More Inference-Efficient Than GLM 5.2 — ChrSzegedy · 2026-10-06
- Reflection AI open-sources Beam: 501B-param MoE claiming 3-4x inference efficiency over GLM-5.2 — rohanpaul_ai · 2026-10-07
24 near-duplicate retellings: bigblueboo · belizgunel · bhutanisanyam1 · mchiang0610 · RebeccaBellan · SumitGup · shakoistsLog · bigblueboo · OfirPress · gethackteam · gethackteam · bigblueboo · alexpolozov · brandondamos · ctnzr · seanmcdonaldxyz · YugeTen · alexpolozov · alexpolozov · bigblueboo · simonguozirui · luke_drago_ · AccBalanced · SergioPaniego