The company has released Beam, a 501-billion-parameter model designed to complete coding, reasoning and agentic tasks with fewer inference tokens. VentureBeat reported that Reflection is positioning the system against Chinese open models that have become central to many developer workflows, including systems valued for their low operating costs and strong coding performance.
Beam is not simply a larger model competing on scale. Its architecture uses a sparse mixture-of-experts design, meaning only part of the model is activated for each token. In its release announcement, Reflection said that Beam’s weights are available on Hugging Face under the Apache 2.0 license, and described the model as built for efficient coding, reasoning and long-horizon agentic work.
Read Reflection AI Bets That Fewer Tokens Can Beat Bigger AI Systems
The official model card lists 501.3 billion total parameters, with 23.4 billion active parameters, and a context window of one million tokens. That context length is intended for tasks in which an agent must hold large amounts of code, documentation or prior interaction in view. Reflection’s broader argument is that a model can be expensive to host in absolute terms while still be economical to use if it reaches a solution with substantially less generated text.
That distinction matters because token pricing is only one part of the bill. Every additional attempt can consume compute, increase latency and create more opportunities for an agent to lose track of its objective. A model that needs half as many tokens to complete a software task could, in theory, reduce the cost of a finished job even if its individual tokens are not cheaper.
“Reflection's chart shows Beam using a third of GLM-5.2's tokens on one test Reflection's October 5 blog post says Beam”
In its introduction to Beam, Reflection said the model was trained for coding and agentic use, and claimed that it could deliver competitive results with fewer inference tokens. The company is therefore asking buyers to judge efficiency as a measure of completed work rather than as a simple measure of model size or per-token price.
The difficulty is that token efficiency is hard to separate from the system surrounding a model. The prompt, tool design, reasoning settings and evaluation harness can all change how many tokens an agent uses and whether it succeeds. A model that looks efficient in one setup may be less impressive when it has to navigate unfamiliar tools or recover from mistakes.
That is the central criticism from independent evaluations. Vals AI’s benchmark results place Beam 21st out of 23 models on its index and report a 4.04 percent result on Terminal-Bench 4.0. Those results suggest to Vals AI that Beam’s performance varies substantially across benchmarks and evaluation setups. Terminal-Bench is especially relevant because it tests whether models can perform practical tasks in a terminal, rather than merely produce an answer in a chat window.
Sesame Disk’s review makes the sharper economic argument. It notes that DeepSeek V4.1 Flash beats Beam on Terminal-Bench while using fewer active parameters per token. If that comparison holds across real workloads, Beam’s large total parameter count and efficiency framing may not translate into lower costs for developers.
DataCamp’s assessment is more measured. DataCamp wrote that Beam is competitive with GLM 5.2 but trails Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash on several agentic coding and reasoning benchmarks. Its conclusion is that efficiency, rather than peak capability, is Beam’s main advantage.
Reflection has also published a Beam safety technical report covering safety training, systemic-risk evaluations, prompt-injection testing and third-party assurance work. Those measures address whether the model can be deployed responsibly, but the report does not independently establish the token-efficiency claim.
That leaves Beam with a narrower, but potentially meaningful, challenge to the market. It does not need to be the strongest model in every benchmark if it can reliably finish common jobs with fewer attempts and less generated text. Yet developers will measure that promise against complete workflows, not laboratory totals. Until Beam’s efficiency survives those tests, “cheaper per task” remains an argument to verify rather than a settled advantage.
Daniel Reyes is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Daniel Reyes's usual length and in Daniel Reyes's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.