TECH & SCIENCE DAILY | AI
NEW YORK
For two years, one question has hung over the Western AI industry like a storm cloud: where is America's answer to DeepSeek? This week, a Brooklyn start-up that has raised billions of dollars without shipping a single flagship model finally stepped into the light.
On Monday, Oct. 5, Reflection AI announced Beam, a 501-billion-parameter open-weight model built for coding, reasoning and the new wave of autonomous 'agentic' software that can plan and carry out tasks on its own. The company says Beam activates just 23 billion parameters for each token it processes, a design that slashes the computing bill, and claims it needs three to four times less inference compute than a leading Chinese rival, Z.ai's GLM-5.2.
The headline numbers are eye-catching. The catch? Almost all of them come from Reflection itself, and the full model weights are not yet in anyone else's hands.
What exactly is Beam?
Beam is what engineers call a sparse mixture-of-experts model. Imagine a giant hospital with hundreds of specialist doctors. Rather than summoning every doctor for every patient, a triage nurse routes each case to the handful who are best suited to it. Beam works the same way: of its 501 billion total parameters, only about 23 billion fire for any given token of text. The result, in theory, is the knowledge of a very large model at the running cost of a much smaller one.
According to the company's launch materials, Beam was trained on 23.8 trillion tokens of curated web data and licensed datasets. Pretraining ran on 6,144 Nvidia GB300 NVL72 GPUs and finished in under four weeks, with what Reflection says was a 92.3 percent goodput rate, a measure of how much of the expensive hardware's time was spent doing useful work rather than sitting idle or recovering from faults. A follow-on reinforcement-learning phase used 10,500 GB300 GPUs over another four weeks and more than 100 million rollouts, the trial-and-error practice runs that teach a model to solve problems and use tools.
Reflection says the model is aimed squarely at 'enterprises, governments, and developers.' In the company's own words, Beam 'combines strong agentic performance, efficient reasoning, and a 500B form factor to give enterprises, governments, and developers a true workhorse open model.'![Reflection AI Beam: 501B-Param Open Model Debuts [2026]](https://tech-insider.org/wp-content/uploads/2026/10/reflection-ai-beam-501b-param-open-model-2026-1.webp)
The scorecard: impressive, but self-graded
Reflection published a set of benchmark results with the launch. On SWE-Bench Verified, a widely watched test of whether an AI can fix real bugs in real software projects, Beam scored 80.9. On Terminal-Bench 2.1, which measures how well a model can operate a command-line computer, it scored 80.1. On AIME 2026, a hard high-school mathematics competition set, it posted 97.8.
The company's main head-to-head comparison was against GLM-5.2 from China's Z.ai, which reportedly scores 81.0 on Terminal-Bench, a hair ahead of Beam. Reflection's pitch is not that Beam beats the best Chinese model on raw performance. It is that Beam gets almost the same results while costing far less to run, which matters enormously to any business that has to pay a bill for every million tokens.
There are important caveats. Every one of these scores is self-reported. No independent lab had verified them at launch. Reflection also named DeepSeek, Alibaba's Qwen, Moonshot AI's Kimi and MiniMax as targets, but did not publish matching head-to-head benchmarks against them. And for now Beam is available only through an early-access waitlist, with the weights expected to be released under the permissive Apache 2.0 license in late October 2026.
That last point is the one to watch. Open-weight AI lives and dies by what the community does after the files land: independent testing, fine-tuning, quantization and the inevitable stress tests. Until then, Beam's scorecard is a promise rather than a proof.
From coding agents to a national mission
Reflection's journey to this week has been anything but conventional. The company was founded in March 2024 by Misha Laskin, a former Gemini researcher at Google DeepMind with a Berkeley postdoc background, and Ioannis Antonoglou, a founding engineer at DeepMind who worked on DQN, AlphaGo, AlphaZero and Gemini's post-training. Both are specialists in using reinforcement learning to sharpen language models.
The company first emerged from stealth in March 2025 with $130 million and a valuation of roughly $545 million, pitching autonomous coding systems. Then it changed course. In October 2025 it raised $2 billion at an $8 billion valuation, a roughly fifteen-fold leap in about seven months, from backers including Nvidia, Sequoia, DST, Lightspeed and GIC. The new mission: to become America's open frontier AI lab.
The logic was laid out by Antonoglou: 'If you want to do reinforcement learning at the frontier, you need an extremely powerful base model to post-train.' And Laskin framed the stakes in blunt, geopolitical terms. 'If we don't do anything about it,' he told TechCrunch at the time, 'then effectively, the global standard of intelligence will be built by someone else.'
He also pointed to an opening in the market: many enterprises and governments, he said, avoid Chinese models because of legal concerns, leaving a gap for a Western open alternative. Reflection's definition of 'open' is also worth noting. It releases model weights for anyone to use and tinker with, while keeping the datasets and the training infrastructure proprietary. As Laskin put it, 'The model weights anyone can use and start tinkering with them. The infrastructure stack, only a select handful of companies can actually use that.'
The $7.3 billion compute bet
Building a model like Beam takes more than brains. It takes a staggering amount of silicon, and Reflection has been quietly assembling it. Reported figures put its compute commitments at about $7.3 billion through 2029: a $6.3 billion deal struck in June 2026 for access to Nvidia GB300 chips at SpaceX's Colossus 2 data center, plus a $1 billion deal in July with cloud provider Nebius.
Those numbers say something about the new economics of AI. A start-up founded just over two years ago is now a customer of the rocket company's computing arm, and is spending sums that once belonged only to the biggest tech giants. Reports have also placed Reflection's valuation at somewhere between $20 billion and $25 billion this year, though those figures should be treated as reported rather than confirmed.
Why the delay mattered
Beam arrives late. Early in 2026 Reflection had promised its first frontier model, trained on 'tens of trillions of tokens,' and as of March commentators were still noting that the model had not shipped, that the company's Asimov coding agent remained on a waitlist, and that it had published no research papers, only product documentation. One analysis flagged the risk of pivoting from a shippable product, coding tools, to competing at the foundation-model layer against rivals with more compute and data.
Seven months on, the company has at least put a concrete, technically detailed model on the table. Whether it lands as a landmark or a footnote will depend on the open release and what independent testers find.
Meanwhile, the competition has not stood still. In just the first days of October, a stream of smaller open and low-cost models appeared: Cloudflare's Clef decision models, Amazon's tiny Strands Decider, Microsoft's new speech models and Bilibili's 35-billion-parameter translation model, among others. The open-model race is crowded, fast and increasingly global.
Expert perspective: the open-weights argument
Supporters of Reflection's approach argue that a healthy open ecosystem is a strategic asset. Open weights let universities, start-ups and governments inspect, adapt and run models on their own hardware, without sending sensitive data to someone else's servers. For regulated industries and public agencies, that control can be the difference between adopting AI and sitting it out.
Skeptics counter that 'open weights' is not the same as fully open source. Because Reflection keeps training data and infrastructure to itself, outsiders cannot reproduce Beam from scratch. And they point out that the business model is unproven; coverage of the company has noted that revenue seems tied heavily to government demand for sovereign AI, rather than to a broad commercial product line.
Both camps agree on one thing: the numbers have to survive contact with the public. Self-reported benchmarks are the opening bid, not the final price.
What this means for you
If you are an ordinary user, you will probably never download a 501-billion-parameter model. But you may well use products built on one. Cheaper-to-run open models tend to push down the prices of AI features inside apps, customer-service bots, coding assistants and workplace tools. If Reflection's efficiency claim holds, it adds to the pressure on every AI provider to cut costs.
If you work in software, Beam's targets are your daily tools: bug-fixing, terminal commands and multi-step agents. An open model that performs near the top of the pack lets companies build custom assistants without being locked to one vendor.
If you care about where AI power sits, Beam is a test of whether the United States can field a credible open competitor to the Chinese labs that have dominated the open-weight conversation. A Western open model with a permissive Apache 2.0 license is, at minimum, a new option for firms that have been wary of Chinese alternatives.
The bottom line
Reflection AI has spent billions, signed a $7.3 billion compute bill and made a very public promise. With Beam, it has finally shipped something to point at. What it has not yet done is prove it.
The next few weeks are the real launch. When the weights drop in late October, the open-source community will do what it always does: pull the model apart, run its own tests and report back. If Beam holds up, America's open-AI moonshot will have found its rocket. If the self-reported scores wobble, it will be another reminder of how hard it is to catch a field that moves this fast.
Either way, the race for open AI just got a lot more interesting.
