Together AI Hosts Reflection's 501B-Parameter Beam Model

Reflection AI‘s first open-weight model, Beam, went live on Together AI‘s inference platform on Friday, built as a roughly 501-billion-parameter mixture-of-experts system trained entirely on Together’s GPU clusters.

Key Takeaways

Beam went live on Together AI’s inference platform on Friday, Oct. 16, as Reflection AI’s first open-weight model
Beam is a roughly 501-billion-parameter mixture-of-experts system trained entirely on Together AI’s GPU clusters
Developers can download, inspect, run and fine-tune Beam on their own infrastructure
Reflection AI has not published Beam’s specific active-parameter figure alongside its total parameter count

Together AI said in a post that Beam represents “America’s next top model,” framing the release as a domestic open-weight entrant in a field increasingly dominated by Chinese labs such as DeepSeek and Alibaba‘s Qwen team.

A Hugging Face blog describing the model calls it “a 501B open model you can own, customize, and scale.”

A mixture-of-experts model splits its parameters into specialized sub-networks, or experts, and activates only a fraction of them per query, letting a system with hundreds of billions of total parameters run at a fraction of the compute cost of a fully dense model the same size.

That activation pattern matters because GPU capacity, power and cost per unit of inference, not headline parameter totals, set the practical limits on deployment.

Beam open-weight model availability means developers can download the weights, inspect them and run or fine-tune the system on their own infrastructure rather than accessing it only through a paid API, the arrangement that governs closed models from OpenAI and Anthropic.

Together AI’s hosting gives Beam immediate commercial inference access alongside the open release, a distribution pattern Together has used before for other open-weight launches.

Why A 501-Billion Parameter Count Matters For Open Models

A parameter count near 501 billion places Beam in the same broad weight class as some of the largest openly released mixture-of-experts systems to date, though the activated parameter count per query, the actual compute load during inference, is typically far smaller than the total.

Reflection AI has not published the specific active-parameter figure alongside the total count.

Also Read: Liquid AI Bets On Privacy With 24B On-Device Model, Skips Benchmark Claims

The company’s pitch rests on training provenance. Beam was trained on Together AI’s own GPU clusters rather than a hyperscaler’s cloud, a detail the companies are using to position the release as evidence that compute capacity outside the three or four dominant cloud providers can still produce frontier-adjacent open models.

Open-weight releases at this scale remain rare because training and serving a 501-billion-parameter system requires sustained access to thousands of high-end GPUs for weeks or months.

The spending on that capacity is the more consequential signal: GPU supply, power availability and inference economics determine whether an open model can be trained and served at scale.

From Stealth Startup To Open-Weight Challenger

Reflection AI emerged from stealth with backing aimed at building large language models domestically, positioning itself against both closed US labs and open Chinese competitors. Beam is the company’s first public model weight release, a milestone that shifts Reflection from a research-stage lab making claims about its training capability into one with a downloadable, independently verifiable artifact.

What Benchmark Scrutiny Comes Next

Independent evaluators will now test Beam against existing open-weight leaders on standard reasoning, coding and knowledge benchmarks, the step that determines whether the 501-billion-parameter count translates into competitive real-world performance.

Together AI’s hosting means enterprise developers can begin running comparative tests within days rather than waiting for a separate inference provider to pick up the weights.