Reflection unveils Beam, an open-weight AI model it pitches as the West's answer to Chinese labs

Beam is a sparse mixture-of-experts model, meaning only part of it runs for each word it produces. That keeps it cheaper to run than its total size suggests.

• Total size: 501 billion parameters.
• Active parameters: 23 billion per token.
• Training data: 23.8 trillion tokens from the web and licensed datasets.
• Context window: 1 million tokens, TechCrunch reports.
• Focus: coding, reasoning and agentic tasks. It's text-only.

For comparison: Z.ai's GLM 5.2 has roughly 744 billion total parameters with 40 billion active, according to TechCrunch.

Not out yet: Beam is still in final red-teaming and evaluations. Reflection says it will release the weights, a technical report, a model card and developer tools later this month. You can sign up for early access now.

Here's Reflection's claim, in short: Beam matches leading Chinese open models on reasoning while using three to four times less inference compute.

Versus China's best open models:

• GLM-5.2 (Z.ai): Beam scores comparably on advanced reasoning benchmarks with 3-4x less compute, Reflection says.
• Qwen 3.8-Max (Alibaba): Beam is "approaching" its performance on coding and agentic tasks. Reflection says efficiency gains are even bigger against huge models like Qwen 3.8-Max, which need much more compute per token.
• Kimi K3: Reflection admits frontier open models like Kimi K3 remain ahead on raw capability. Beam's edge is efficiency.

Some sample scores from Reflection's own tables:

• SWE-bench Verified: 80.9, versus 77.6 for Thinking Machines Lab's Inkling.
• Terminal Bench v2.1: 80.1, close to GLM 5.2's 81.0.
• GPQA Diamond: 90.5, versus 91.2 for GLM 5.2.
• AIME 2026: 97.8.

A big caveat: these claims haven't been independently verified, TechCrunch notes.

Reflection leaned heavily on reinforcement learning, where a model improves by trying tasks and getting feedback.

• The hardware: 10,500 Nvidia GB300 GPUs running for four weeks.
• The scale: more than 100 million rollouts, or practice attempts, with up to 256,000 tokens of context each, using about 1.3 billion sandboxes.
• The tasks: one million coding, agentic and STEM training environments.

Reflection believes this is one of the largest reinforcement learning runs by any open lab so far. Its capabilities kept improving with more compute, "with no sign of a plateau," the company says. For comparison, it says Inkling was trained on 30 million rollouts.

Adjustable thinking: users can set a "reasoning effort" level. Lower settings give shorter answers, and higher settings let Beam think longer on hard tasks.

Reflection shared several demos to show Beam's range.

• A live NYC subway map: Beam searched for documentation, worked out authentication, found map data and built a live-updating subway map from public MTA data, frontend and backend included.
• A 3D game: asked to build a p5.js game where an astronaut falls toward Earth dodging asteroids, Beam reasoned about how the visuals should look in text, despite not being multimodal.
• Training another model: plugged into the OpenCode tool, Beam read documentation and wrote a notebook to fine-tune Google's smallest Gemma-4 model on a text-to-SQL task. Reflection says that raised Gemma's accuracy on held-out tests by 66.5%.
• A brand-new puzzle: on a viral land-or-water map puzzle only a few days old, Beam got 95.5% right, Reflection says.

Skills it wasn't taught: during training on reasoning, coding and terminal tasks, Beam also got better at web browsing, even though browsing wasn't in its training mix, Reflection says. Given web access, it learned on its own to query other AI models and use text-recognition tools to read documents.

Data quality mattered: Reflection says cutting corners on training task quality led to plateaus. Its team built nearly one million tasks, mostly through synthetic data pipelines, and filtered out ones that were too easy, impossible, guessable or broken.

Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both former Google DeepMind researchers.

• Funding: about $4.7 billion raised from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners, per PitchBook data cited by TechCrunch. Semafor adds Citigroup as an investor.
• Valuation: its latest round, closed in June, valued it at $25 billion before the new money.
• Compute: this summer, it signed deals worth more than $7 billion with SpaceX and Nebius for access to Nvidia's GB300 chips through 2029, TechCrunch reports.

It's already training its next model, which Laskin told Semafor would be "much more" powerful than Beam.

Reflection is aiming Beam at enterprises, governments and developers.

"AI factories": the pitch is to let institutions build their own customized, local AI systems by training Reflection's models on their own data, TechCrunch reports. Nvidia CEO Jensen Huang has long championed the idea, which would also boost demand for Nvidia GPUs.

Early interest: Axios reported that hedge funds and trading firms are among those keen on such systems. Reflection has also begun testing a sovereign AI factory partnership with South Korea's Shinsegae Group, according to TechCrunch.

Distribution: Reflection says Beam will be available through hyperscalers and smaller cloud providers, with integrations across open-source libraries at launch.

Chinese companies have led the world in open-source AI, Semafor notes. Several large Western companies have adopted Chinese models to cut costs on tasks like customer service and routine coding.

The security worry: because open models can be downloaded and run with fewer safeguards, their rise has raised concerns. US and UK government evaluators found recent Chinese models could help hackers exploit code vulnerabilities, according to Semafor.

The competition: Beam joins other Western open-weight models from Thinking Machines Lab, Mistral, Meta and Cohere.

Safety testing: Laskin told Semafor that Reflection is working with the US Center for Advancing Innovation and Standards for Super Intelligence and the UK's AI Safety Institute to assess Beam. He also wants open-model builders to have a voice in Washington's AI debate.

"You want multiple voices around the table, both open and closed," he said.

The political backdrop: Beam arrives as calls for AI regulation grow louder in the US after a string of alarming security incidents, Semafor notes. President Trump has pushed for companies to regulate themselves rather than face federal oversight, but that could change quickly if Democrats win Congress in November, according to Semafor.

• Developers: a strong, efficient open model for coding and agents could arrive this month. Watch for independent benchmark results once the weights are out.
• Businesses and governments: Beam offers a possible alternative to Chinese open models for systems you run and control yourself.
• Costs: if the 3-4x efficiency claim holds, running capable open models could get noticeably cheaper.