Reflection AI Launches Beam, a 501B Open-Weight Coding Model

Reflection AI, a startup built by two former Google DeepMind researchers, just put a half-trillion-parameter AI model on the internet for anyone to download — no subscription, no API key, no usage cap.

What's actually in the release

The model is called Beam, and the numbers are worth sitting with. It has 501 billion parameters in total, but it's built as a mixture-of-experts (MoE) system — a design where the model is split into many smaller specialist networks, and only a fraction of them switch on for any given request. For Beam, that means roughly 23 billion parameters actually do the work on a typical query, which is why Reflection says it runs far cheaper than its headline size suggests. It also ships with an effective context window of around 1 million tokens — enough to hold a fairly large codebase in its "memory" at once — and comes in three numerical precision formats (BF16, FP8 and NVFP4), so engineers can trade a bit of accuracy for speed and lower memory use depending on their hardware.

Why the licence is the real headline

What makes this more than just another model drop is how Reflection is giving it away. Beam's weights are published on Hugging Face under Apache 2.0, a permissive open-source licence that lets anyone download the model, run it on their own servers, fine-tune it, and even build and sell products on top of it — without paying Reflection anything or asking permission first. That's a real departure from how most frontier-grade AI still works: you rent access to a closed model through an API, your prompts and code pass through someone else's servers, and you pay per token for the privilege. Reflection also open-sourced Mirror CLI, the Rust-based command-line agent framework it used to train and test Beam, and says the model launched alongside more than 40 partner integrations spanning cloud providers and developer-tool makers.

• 501 billion total parameters, ~23 billion active per query (mixture-of-experts)
• Roughly 1 million token effective context window
• Apache 2.0 licence — free to self-host, modify and resell
• Purpose-built for agentic coding: fixing GitHub issues, multi-step tool use, long-running workflows
• Ships in BF16, FP8 and NVFP4 checkpoint formats

An American entry in a race China's been winning

Reflection has been fairly open about why Beam exists. Giving away frontier-grade model weights has been mostly a Chinese story this year — DeepSeek, Moonshot's Kimi and Zhipu's GLM line have all shipped powerful open models while most big American labs kept theirs sealed behind paid APIs. Backed by Nvidia and founded by Misha Laskin and Ioannis Antonoglou, both former staff research scientists at Google DeepMind, Reflection is explicitly pitching Beam as a US-built entry in that same open lane, aimed squarely at coding and the kind of multi-step "agent" work — where a model plans, acts, checks its own output and tries again — that's become the industry's biggest obsession.

The launch hasn't gone entirely unchallenged, though. Reflection's own charts show Beam beating GLM-5.2, a model with roughly 250 billion more parameters, while using a third to a quarter of the compute at inference. But independent reviewers have pointed out that on the two benchmarks that matter most for Beam's actual job — Terminal-Bench and DeepSWE, both tests of how well a model handles real agentic coding tasks — it trails DeepSeek's V4.1-Flash, a model we've covered before for slashing the memory costs of long-running AI agents, by a noticeable margin on both.

A free model is only as useful as what it can actually do — and on the exact job it was built for, Beam is competing with rivals it hasn't yet beaten.

What it means for developers in India

For an Indian startup building an AI coding assistant or an internal dev-ops agent, the appeal of something like Beam isn't just that it's free — it's where the model runs. Sending proprietary code or customer data to a foreign API for every request is exactly the kind of data flow that India's Digital Personal Data Protection (DPDP) Act was written to make companies think twice about. A model you can download and host yourself, on an Indian cloud provider or even in your own data centre, sidesteps that question entirely, along with the rupee-dollar exposure that comes from paying per-token API bills in a currency that keeps moving against you. It also plugs neatly into the government's own IndiaAI Mission, which has been pushing domestic firms toward building on or around open-weight foundation models rather than renting intelligence from Silicon Valley. Whether Beam specifically becomes the model Indian teams reach for will depend on those benchmark gaps closing — but the fact that a credible, frontier-scale alternative now exists outside the usual closed-API lineup is the part worth paying attention to.

The real test for Beam won't be this week's benchmark charts, it'll be whether developers actually build on it six months from now, after the novelty of a free 501-billion-parameter download wears off. Open weights lower the barrier to entry; they don't guarantee anyone walks through it.