Reflection AI Launches Open-Weight Model ‘Beam,’ Pushing the Western Open Frontier
Open-weight large language models have been China’s game for the past year. DeepSeek shook up the industry with low-cost, high-performance models, and Moonshot AI’s Kimi line, Alibaba’s Qwen, and Zhipu AI’s GLM have been releasing weights and trading places at the top of the leaderboard ever since. Meanwhile, America’s leading labs — OpenAI, Google, Anthropic — keep their best models closed. Into that gap, often described as the “Western open-weight void,” steps not a Big Tech giant but a startup founded in 2024 by two Google DeepMind alumni.

Reflection AI unveiled its first open-weight model, Beam. It’s a sparse mixture-of-experts model with 501 billion total parameters, only 23 billion of which are active at inference, built for coding, reasoning, and agentic work. The company said the release “advances the Western open-weight frontier.” Final red-teaming and evaluations are still underway, so the weights themselves will ship under an Apache 2.0 license later this month; for now, the company is only taking early-access signups.
Two DeepMind Alumni, From Coding Agents to Open-Weight Models
Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou. Laskin, now CEO, led reward modeling for Gemini at Google DeepMind, while Antonoglou, the CTO, co-created AlphaGo and AlphaZero and ran RLHF (reinforcement learning from human feedback) for Gemini. The company’s original pitch was “superintelligent autonomous coding agents,” but it has since widened its ambitions into an “open frontier AI lab” that builds and releases its own foundation models.
Its fundraising kept pace with that pivot. It exited stealth in March 2025 with a combined seed and Series A of $130 million, then raised a $2 billion Series B led by NVIDIA that October at an $8 billion valuation. According to CNBC, it raised a further $2.5 billion Series C this year, pushing its valuation to $25 billion — nearly a 50x jump in two years.
23.8 Trillion Tokens of Pretraining, Then Four Weeks of RL on 10,500 GPUs
Reflection AI says Beam’s edge comes from heavy investment in both pretraining and reinforcement learning. It was pretrained on 23.8 trillion curated tokens from the web and proprietary licensed datasets, matching or beating similarly sized open base models. It then went through a high-compute RL run on 10,500 NVIDIA GB300 GPUs over four weeks, generating more than 100 million rollouts. Where other frontier labs have only disclosed generation-over-generation RL-compute multipliers, Reflection published specific GPU counts and rollout numbers.
On benchmarks, the company says Beam matches or beats Zhipu AI’s GLM 5.2 and Alibaba’s Qwen 3.8-Max, and approaches larger models like Moonshot AI’s Kimi K3 on coding and agentic tasks — though it’s worth noting these are the company’s own measurements, not yet independently verified since the weights aren’t public. What Reflection emphasizes most isn’t raw capability but inference efficiency: it claims performance on par with GLM 5.2 using 3-4x less inference compute.
Competitors: China Owns Open Weights, the West’s Bet Is on New Labs
Chinese companies have effectively dominated open-weight releases over the past year. DeepSeek recently shipped V4.1 Flash, Zhipu AI (Z.ai) released GLM 5.3, and Moonshot AI put out Kimi K3, a 2.8-trillion-parameter model that’s the largest open-weight release to date. Alibaba bills Qwen 3.8-Max as its “first open-sourced Max-class model,” though there’s some outside disagreement over whether the released checkpoint fully matches the flagship running in production. On the US side, NVIDIA open-sourced Nemotron 3 Ultra, but leaders like OpenAI, Google, and Anthropic keep their top models closed.
The company Reflection AI gets compared to most is Thinking Machines Lab, which is on a similar “open frontier” path. Founded by former OpenAI CTO Mira Murati, it released its own first open-weight model, Inkling, in July at 975 billion parameters, and is reportedly raising more funding at a $40 billion valuation. Both companies were founded by Big Tech alumni, raised large rounds, and ship their own open-weight models — but where Thinking Machines leans multimodal, Reflection is betting on coding and agentic performance plus inference efficiency as its edge.
Zoom out further and both sit inside the broader “neolab” category of research-first AI startups. Ilya Sutskever’s Safe Superintelligence (SSI) holds a $32 billion valuation with no shipped product; Fei-Fei Li’s World Labs was just acquired by AMD for $8.2 billion; and Yann LeCun’s AMI is betting on world models as the successor to LLMs. Each has taken a different technical path, and Reflection AI is staking out the most immediately practical one: an open-weight coding and agentic model developers can actually use today.
MORE FROM THE POST
- Mistral Unveils 1-Trillion-Parameter ‘Large 4,’ Europe’s Strongest Open-Weight Model
- Thinking Machines Lab Unveils Inkling 2 Small as It Reportedly Nears a $5B Round
- Meta Releases Open-Weight ‘Muse Spark Open,’ Optimized for Apple Mac Studio
- NVIDIA Plans Investments in Korean AI Startups Motif, Upstage and RLWRLD
- Meta Unveils Muse Wonder and Wonder Full Agent Models
Share
Most Read
- 1
- 2
- 3
- 4
- 5


Leave a Reply