Reflection AI today released the weights of Beam, its first open model, and is pitching it against the Chinese open models that many U.S. companies are already embracing because of their lower costs.
Reflection is betting that Beam, a 501-billion-parameter model, can finish a task in far fewer tokens than other models its size, which Reflection says lowers the cost of each finished task.
Beam arrives a day after Meta released Muse Spark Open, an 816-billion-parameter model that Meta calls the most capable U.S. open-weight model to date. Meta pitches its bigger model on capability. Neither Meta's model nor Reflection's matches the performance of the top Chinese models on Terminal-Bench 2.1, a widely used test of agent tasks on a command line, according to each company's own published scores. But Reflection is betting that as long as it offers competitive performance for such tasks, enterprises will care more about how many tokens a model needs to finish each task. Beam's efficiency delivers "strong model capabilities at lower cost," Reflection wrote in its October 5 blog post, where it first previewed the details of its model.
In that post, Reflection said it trained the model with intensive reinforcement learning, the training stage in which a model practices tasks and is rewarded for results – and that the company had placed a penalty on the model whenever it used unnecessary tokens.
In an exclusive interview with VentureBeat, Reflection co-founder and Chief Technology Officer Ioannis Antonoglou explained why that penalty is hard to get right. Asked what Beam gave up to produce shorter, more efficient answers, Antonoglou said, "You don't give up anything. You just penalize it in terms of how many tokens it can use."
The hard part, Antonoglou said, is that a team that applies the penalty “naively” ends up with a model that "will just learn to not solve the task," because producing fewer tokens is easier than doing the work. To keep Beam from giving up on tasks, he said, Reflection ran a very large amount of reinforcement learning. With enough reinforcement learning, Antonoglou said, a model learns to solve the task and stay brief at the same time.
Reflection AI conducted what it describes as one of the largest reinforcement learning runs ever undertaken by an open-weight lab. Over a four-week period, the team ran more than 100 million "rollouts" – essentially complete, trial-and-error problem-solving attempts where the model tests its reasoning and learns from its mistakes – powered by a cluster of 10,500 of Nvidia’s cutting-edge GB300 GPUs.
Counting evaluations and experiments, Reflection ran about 1.3 billion sandboxes, the isolated computing environments in which the model practices tasks, and its computing fleet peaked at 883,006 virtual CPUs across 23 clusters on Google Cloud and Amazon Web Services, according to the company’s launch post this morning.
Of all of the Chinese labs, DeepSeek in particular has been praised for its approach to optimization and efficiency. Antonoglou, however, said Beam compares favorably with DeepSeek’s latest V4.1 Flash model. He cited Artificial Analysis, the independent benchmarking firm, which said that early indicators suggest Beam will be “one of the most token-efficient open models we’ve seen for its level of intelligence.”
However, DeepSeek does not appear in the chart where Reflection compares Beam's scores with the number of tokens each model generates. That chart, in the October 5 blog post, covers Z.ai's GLM-5.2, two versions of Alibaba's Qwen 3.8, Thinking Machines Lab's Inkling, Nvidia's Nemotron 3 Ultra and Meta's earlier, smaller open model, Muse Glimmer.
Why a model with pricier tokens can cost less per task
A company's AI bill is the price of each token multiplied by the number of tokens a task takes, but most price comparisons show only the price per token.
Take two hypothetical models. Model A charges $1 per million tokens and needs 30,000 tokens to fix a bug, so the fix costs 3 cents. Model B charges twice as much, $2 per million tokens, but needs only 10,000 tokens, so Model B finishes the same fix for 2 cents. Providers often charge different prices for the tokens a model reads and the tokens a model writes, but the same multiplication applies to each.
The price per token is the number buyers see on a price sheet, and a growing share of the respondents to VentureBeat's monthly survey on AI infrastructure name it as a factor. In the September survey, 23% of respondents named cost per million tokens as a factor in choosing AI infrastructure, up from 11% in August.
When a model generates many tokens to solve a task, the end customer is the one paying the cost of that, Antonoglou said.
The models generating the most tokens on OpenRouter, the service that routes developers' requests to many models, are models of Beam's size. He called them the workhorse, or "flash" models – the fast, mid-sized models that labs such as DeepSeek release for everyday work. Beam has 501 billion parameters in total and uses 23 billion of them, its active parameters, for each token. Beam aims to be a workhorse model for coding and agent work, according to the company.
Meta's Muse Spark Open uses nearly twice Beam's active parameters per token
By comparison, Meta's new open model, Muse Spark Open, uses 42 billion active parameters for each token, against Beam's 23 billion. Reflection estimates the compute behind each token from a model's active parameters, so by VentureBeat's calculation using that method, each Muse Spark Open token takes about 1.8 times the compute of a Beam token.
Meta has not published how many tokens Muse Spark Open uses per task. Hugging Face's launch post says Muse Spark Open's reasoning effort defaults to the highest setting, and Hugging Face advises developers to set it lower to use fewer tokens on easy tasks.
Meta's model had a price on its first day, yesterday: Four providers on OpenRouter list Muse Spark Open at $1.25 per million input tokens and $4.25 per million output tokens. Reflection does not yet offer pricing, however. It said it is still setting the price for its own hosted service, which is free during during a public beta, and that partners such as Baseten set their own prices.
The two licenses also differ. Beam's weights use the Apache 2.0 license. Meta's Muse Model License requires products built on the model to display "Built with Muse Spark Open." Meta also requires a separate agreement from any company with more than $20 million in revenue that sells access to the model as a service.
Reflection said it has not compared Beam directly with Muse Spark Open. But Reflection said Beam's reasoning efficiency sets Beam apart, along with what Reflection calls an "open operating system," the software to run and deploy Beam that Reflection is launching with partners. It welcomed the competition from Meta, saying labs releasing several open models at once give companies and developers more choice.
The two launches show how U.S. labs are answering China’s lead in open models. Chinese labs still hold the top of the open-model performance rankings, but Meta is answering with power: a larger model that Meta says beats the leading Chinese models on two agent and tool-use tests, JobBench and MCP-Atlas. Reflection is answering with leanness: a smaller model trained to finish each task in fewer tokens. For companies choosing an open model, the two launches split one question into two. One is which model does the most. The other is which model does the job for the least money.
DoorDash, Coinbase and Pinterest credit open models, including Chinese ones, with lower costs
Several well-known U.S. companies have credited open models, many of them Chinese, with lowering their costs. Pinterest built Navigator 1, the conversational model behind its shopping assistant, on Alibaba's Qwen. "We have this conversational AI model that we've recently rolled out called Navigator 1 that's built on Qwen 3 VL," Pinterest Chief Technology Officer Matt Madrigal told VentureBeat on the Beyond the Pilot podcast. Pinterest CEO Bill Ready told investors in November 2025 that the company was seeing a 90% reduction in cost using fine-tuned open models versus leading off-the-shelf proprietary models.
DoorDash Chief Technology Officer Andy Fang credited Moonshot's Kimi with "better quality [at] cheaper cost" for code review, according to Fortune. Coinbase CEO Brian Armstrong said the company cut AI spending nearly in half after making Z.ai's GLM and Moonshot's Kimi the default models in its internal tools, according to MLQ. Airbnb CEO Brian Chesky called Qwen "fast and cheap" in October 2025, according to Bloomberg.
At Vercel, which routes developers' requests to many AI models, DeepSeek's share of the tokens Vercel handled rose from under 1% to 17% in May 2026, according to Rest of World. DeepSeek's share of the money developers spent on models through Vercel stayed near 1%. DeepSeek handled far more of the work, measured in tokens, than its share of the spending, because each DeepSeek token is so cheap.
Antonoglou did not dispute that U.S. companies are adopting Chinese models. He said Beam will "win on merit rather than on anything else."
Antonoglou also named the risks he sees in running a Chinese model on a company's own servers.
Regulation around the use of Chinese models could change, he said, and companies face questions of liability and of provenance, meaning where a model came from and how it was trained. Adopting any model also makes a company dependent on the lab behind it, he said, so widespread use of Chinese models makes “core pieces of the U.S. economy” dependent on Chinese technology. The newer risk comes from agents. Models no longer just chat. They act as agents with access to a company’s tools and sensitive data. A company cannot be sure, Antonoglou said, that a model was not “trained to just activate certain behaviors in certain scenarios,” meaning a hidden behavior that switches on only under particular conditions. He added: "This is not to say that Chinese models are bad."
Many companies running open models may not know whether they carry that provenance risk. In VentureBeat's September survey, 32 respondents said they run open-weight models. Of those 32 respondents, 56% could not say with confidence where the models in their stack originate, or did not know whether any came from a Chinese lab. Another 31% said none of their models came from a Chinese lab. The remaining 13% said Chinese models came into their stack through a vendor, platform or framework. No respondent picked the answer that they had chosen a Chinese model deliberately.
Cursor shows how a Chinese model can arrive unannounced. Cursor, the maker of an AI coding tool, built its Composer 2 model on Moonshot's Kimi K2.5. Cursor confirmed that Composer 2 started from Kimi K2.5 only after a developer found the model name in Cursor's traffic, as VentureBeat reported in March.
VentureBeat asked Reflection whether Beam learned from the outputs of other companies' models, Chinese or otherwise. Reflection said it trained Beam from scratch, starting from new weights, and that it does not train on any other model's outputs. Reflection said the synthetic data its October 5 blog post describes means the training environments Reflection built for reinforcement learning.
Reflection penalized Beam for extra tokens during four weeks of reinforcement learning
Reflection's reinforcement learning run trained the model for coding, agent, science and math work. Reflection wrote that Beam's scores rose early in the run even as its answers got shorter. Users can set a reasoning effort level, and lower settings produce shorter answers.
Antonoglou explained the trade-off that Reflection balances with the penalty. Bigger models need fewer tokens to solve a task, but each token costs more to compute. Smaller models are cheaper per token but need more tokens. "What you want to achieve is a model that for its model size uses the minimum number of tokens to achieve the task," he said.
Reflection's team also had to stop Beam from cheating. In reward hacking, a model finds a shortcut that earns the training score without doing the work. Antonoglou described catching reward hacking as a cat-and-mouse game: The team patches one shortcut, and the model finds another.
Not every shortcut was a problem. Beam reads only text, so when it was given web access and needed to read a picture of text, it went looking for outside help. Beam taught itself to search for and query other AI models online, and to use OCR services, which convert a picture of text into text, the company said. Antonoglou called Beam’s request for help a good kind of shortcut, the model “getting creative in order to solve the task,” not doing something malicious."
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 "achieves scores comparable to GLM-5.2 while using 3-4x less inference compute" on advanced reasoning tests. Reflection’s compute figure is an estimate, not a measurement. It counts the work a model does to write each token, which grows with the model’s active parameters, but it leaves out the work of reading the prompt. That missing piece matters most in agent and coding tasks, where prompts are often long.
Reflection Vice President of Product Joe Spisak said developers testing Beam have seen similar gains. He said developers in Beam’s private preview, which opened to a waitlist on October 5, have seen “almost like 3x more efficient reasoning compared to a lot of other models.”
The token counts in Reflection's chart vary by test. On Terminal-Bench 2.1, Beam scores about 80% while generating roughly 25,000 tokens per attempt, by VentureBeat's reading of the chart. GLM-5.2 scores slightly lower while generating roughly 80,000 tokens, so Beam uses about a third as many tokens on that test. On DeepSWE, a software engineering test, the two models reach about the same score, and Beam uses a little more than half as many tokens as GLM-5.2.
Beam's smaller size adds to the token savings in Reflection's compute estimate. Reflection told VentureBeat it counted GLM-5.2 at 40 billion active parameters per token, against Beam's 23 billion, so Beam needs a little more than half the compute for each token on top of generating fewer tokens.
Reflection has not yet put a dollar figure on Beam's efficiency. Asked what one finished coding task costs on Beam compared with DeepSeek or GLM, Spisak said, "We don't really have anything concrete to share there." Baseten is serving Beam through OpenRouter, starting today, Reflection said, but had not listed a price at publication time.
Spisak also conceded that lower token and compute counts may not translate directly into a lower bill. "A DeepSeek model or Qwen model may have better kernels or fused kernels, and they just might be much more efficient," he said, referring to the low-level software that runs a model on specific chips. Reflection said it has worked with partners to build fused kernels for Beam to lower latency and raise throughput, and it is releasing a smaller draft model that speeds up how fast Beam generates tokens.
In Reflection's coding table, DeepSeek V4.1 Flash scores 90.6% on Terminal-Bench 2.1, against Beam's 80.1%. Meta's own model card lists Muse Spark Open at 76.4% on the same test, against 85.0% for Moonshot's Kimi K3.
Antonoglou said Beam is not built to replace frontier models such as Anthropic's Claude or OpenAI's GPT-6 Astra "on everything." Beam targets the tasks that "consume a big chunk of your token budget," he said.
Spisak's lesson from Meta's Llama: Make the model easy to deploy
Spisak ran product on Meta's Llama models 1 through 3. He said Llama 4 went wrong because Meta became too "focused, I would say, on benchmarks and these kind of things and less about the community" of developers building around the model.
Reflection built its open operating system for AI agents with 33 named partners, and Reflection said about nine more partners are awaiting final approval. The software covers running, customizing and deploying Beam. Reflection argues that many labs publish open models on Hugging Face and leave developers to assemble the surrounding software themselves. "We want to make it so easy to develop with, so easy to customize, so easy to deploy," Spisak said.
Reflection is releasing that software under the Apache 2.0 license, the same license as Beam's weights. The software includes containers that let a company pull Beam onto its own servers with one command. "Like boom, you have Beam running on your infrastructure on Kubernetes in essentially a few minutes," Spisak said.
The software also includes Mirror CLI, an open-source coding agent that runs from the command line, written in Rust and Python, which Spisak said Reflection already uses internally to write and review code for its own model.
Spisak said demand so far is strongest for coding agents that companies own and run themselves, including in air-gapped environments cut off from the internet. Dell and HPE will install Beam on companies' own hardware, and Reflection will send its own forward-deployed engineers, staff who work on site with a customer, to help build those coding agents.
Spisak said Beam fits on a single server of eight GPUs. Reflection said the eight-GPU figure applies to Beam's 8-bit and 4-bit versions, which are the formats most commonly used in production. Those versions can be served on eight-GPU servers built with Nvidia's H100, H200 or GB300 chips. The full-precision version, with each number stored in 16 bits, needs "tens of GPUs," Reflection said.
Spisak also said vLLM, the open-source software many companies use to serve models, will publish a blog post today showing Beam's weights running in MXFP4, a 4-bit format that AMD chips and Nvidia's newer Blackwell chips support. Meta showed Muse Spark Open running on two linked Apple Mac Studios, and Hugging Face lists Muse Spark Open's 4-bit version on eight-GPU Nvidia servers.
Reflection has not named a paying customer, but it named one early user. Oura, the maker of the Oura smart ring, tested Beam in early access through Modal, a cloud platform for running AI models that partners with Reflection. “We found that Beam had fewer context leaks and was much more grounded in its answers than other frontier models we evaluated,” Oura’s Health Intelligence Team said in a statement provided by Reflection.
Spisak said the end customers of Reflection’s other partners are "pretty big names. We can't disclose them yet," and that some will become public soon. Reflection's one named partner project is with Shinsegae Group, the South Korean retail conglomerate, which is building what Spisak called an "AI factory" in Korea to serve local users. Spisak said Shinsegae and Reflection are developing shopping agents together, with deployment planned.
Reflection's technology chief calls the Nvidia acquisition report "baseless rumors"
The Financial Times reported on October 10 that Nvidia is in early talks to buy Reflection or invest more. Antonoglou called Nvidia “a close partner and investor in the company" that has supported Reflection since its Series B round. "We remain committed to our mission of building frontier open intelligence as an independent company," he said, calling the report "baseless rumors." He did not say whether talks are taking place.
Antonoglou says deploying Beam safely takes expertise
On safety, Antonoglou said, "I don't think that there's actually a meaningful risk in terms of deploying these systems in an enterprise setting," as long as a company gives the model a clear task and the right tools. That takes expertise, he said: "If you naively try to do it, then you might actually hurt yourself." Reflection published a Beam Safety Report today describing its safety methods and evaluations, and said that outside groups including Gray Swan, Irregular and SecureBio tested Beam before release.
VentureBeat is testing cost per finished task
VentureBeat is testing Beam, Muse Spark Open, DeepSeek V4.1 Flash and GLM-5.2 on what a finished task actually costs, and will publish the results.
VentureBeat's September AI Infrastructure and Compute survey drew 102 qualified respondents representing organizations with 100 or more employees. The question on Chinese-origin models went to the 32 respondents who run open-weight models.
VentureBeat's Agentic Infrastructure Stack Summit, October 28 in San Francisco, takes on inference economics, including how teams do more with less memory and fewer tokens. Attendance is by application: venturebeat.com/agentic-infrastructure-stack-summit