Reflection Releases Beam, a 501B-Parameter Open Model, Free on Ollama
Reflection has released Beam, its first open-weights model built for coding and long agentic tasks, and Ollama, a service that runs models with a single command, is offering it free of charge for now. The 501B figure is the total parameter count, but only 23B are active for each token, much like a few specialists answering a question instead of an entire firm. Beam has a 1M-token context window and ships under the Apache 2.0 license. That means you can test a coding agent on Beam without downloading the weights.
AI-processed from Ollama Blog; edited by Hamidun News
Open models for programming usually fall short in one of two ways: either they are too weak for long tasks, or they are so hard to run that they never get past the demo stage. Reflection has released its first open-weights model, Beam, and arranged with Ollama, a service that lets you run models from the terminal with a single command, to make it available on launch day. For now, you can try Beam for free and without your own hardware.
The 501B in the headline is the model's total parameter count, meaning its entire volume of "knowledge." But only about 23B of those parameters work on each token (a small piece of text), so each step requires noticeably less computation than a conventional model of that size.
What the model is
Beam is built on a mixture-of-experts (MoE) architecture. Picture a large consulting firm: for each question, it doesn't bring in every employee, only a few suitable specialists. The same logic applies here. The model has roughly 501B parameters, and about 23B are activated for each token. It holds a lot of knowledge, yet requires less computation per step than a dense model of the same size.
Beam is aimed at practical use: software development, reasoning, and long agentic chains. The model works on a text-in, text-out basis, with no images or audio. Reflection is betting on scenarios where code, tool calls, logs, documents, and structured responses are central. That means not "ask a question, get a reply," but "write and maintain code, call tools step by step, and reach a goal through dozens of actions."
For such tasks, context matters: the amount of text the model can hold in mind at once. Beam's context is 1M tokens, enough for a large repository, a long history of agent actions, and execution logs. The Apache 2.0 license allows you to study, fine-tune, and embed the model in your own products, including commercial ones, without separate approvals.
How to try it in Ollama
Ollama has launched Beam as a cloud model. You don't need to download hundreds of gigabytes of weights or deal with inference, meaning the process of actually running the model. A single command is enough. The announcement gives three examples:
- ollama launch claude --model beam:cloud
- ollama launch codex --model beam:cloud
- ollama launch opencode --model beam:cloud
In other words, you can plug Beam into familiar coding agents: Claude Code, Codex, or opencode. It's a convenient way to compare the model on your own tasks without rewriting your workflow.
There is also a privacy caveat. According to Ollama and Reflection, the Beam endpoint operates with full zero data retention (ZDR): requests are not stored by the provider. For teams that don't want to keep their code in the cloud, this is an important argument, though such promises should still be checked against the documentation.
What to keep in mind
First, the free access is temporary, and the announcement gives no specific end date. Don't build a production budget around it; you will need to check the follow-on terms with Ollama.
Second, the Ollama announcement contains no benchmarks and no comparisons with other open models. Claims about quality in programming, math, and science remain the developer's claims for now. The details are in Reflection's announcement and on the model card, and that is where you should look for test results.
Third, 501B parameters is a very large model for self-hosting. Even though only 23B are active per token, all the weights must still be held in memory. In practice, most teams will use Beam through the cloud first, and local deployment will be a question for those with a serious cluster.
What developers should do
If you are building coding agents or internal tools, the most sensible move right now is to take a couple of your real tasks and run them through Beam in the same agents you already use. It's worth testing long tasks, where other open models lose the thread, and work with large repositories using the million-token context. If the results are good, the Apache 2.0 license leaves room to fine-tune the model for your stack and avoid depending on a closed API. If they aren't, you will have spent a few minutes and nothing else.
- Nov 21, 2026Author Makes a Narrative Video on Home LTX 2.3 (22B Parameters) Without Subscriptions
- Nov 21, 2026Helion from PyTorch Picks GPU Kernel Settings at Compile Time Instead of 25 Hours of Autotuning
- Nov 21, 2026Anthropic Shows How It Verifies Code That Claude Code Writes 90% Of
- Nov 21, 2026Linger Runs 12 AI Models on Your PC: Up to 5x Faster and $0 in Token Costs
Key Questions Answered
What the model is?
Beam is built on a mixture-of-experts (MoE) architecture. Picture a large consulting firm: for each question, it doesn't bring in every employee, only a few suitable specialists. The same logic applies here. The model has roughly 501B parameters, and about 23B are activated for each token. It holds a lot of knowledge, yet requires less computation per step than a dense model of the same size. Beam is aimed at practical use: software development, reasoning, and long agentic chains.
How to try it in Ollama?
Ollama has launched Beam as a cloud model. You don't need to download hundreds of gigabytes of weights or deal with inference, meaning the process of actually running the model. A single command is enough. The announcement gives three examples: ollama launch claude --model beam:cloud ollama launch codex --model beam:cloud ollama launch opencode --model beam:cloud In other words, you can plug Beam into familiar coding agents: Claude Code, Codex, or opencode.
What to keep in mind?
First, the free access is temporary, and the announcement gives no specific end date. Don't build a production budget around it; you will need to check the follow-on terms with Ollama. Second, the Ollama announcement contains no benchmarks and no comparisons with other open models. Claims about quality in programming, math, and science remain the developer's claims for now. The details are in Reflection's announcement and on the model card, and that is where you should look for test results.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.