Reflection's Beam: A 501B Open-Weight Model That Costs Less to Run, If You Believe the Numbers

Reflection's Beam: A 501B Open-Weight Model That Costs Less to Run, If You Believe the Numbers

The Machine Drifted In

The thing about a new AI model is that it always arrives like a stranger walking into a bar at 2 a.m. Nobody asked for it. Nobody knows what it wants. It just sits down, orders something, and starts talking.

Reflection AI dropped Beam on a Monday. October 5, 2026. A 501-billion-parameter mixture-of-experts model. Twenty-three billion of those parameters wake up for any given task. The rest sleep it off. That's 4.6 percent of the machine doing the work. The rest is dead weight until called upon.

They trained it on 23.8 trillion tokens. The context window stretches to one million tokens. It reads text. It writes text. It does coding and agentic work and reasoning. It is, in their words, a "workhorse."

The weights aren't out yet. They'll come later this month under an Apache 2.0 license. For now you can sign up for a waitlist and poke at an early version through an OpenAI-compatible endpoint. No price published anywhere. That's the first thing worth noticing.

What the Numbers Say When You Line Them Up

Reflection published its own benchmark table. That's the only table there is. No independent verification yet. Artificial Analysis said evaluation was under way and that early indicators looked promising. That's as close to a third-party nod as you get right now.

Here's the table, stripped of the marketing glaze:

Beam wins on DeepSWE against GLM-5.2 by four-tenths of a point. It loses to Kimi K3 on everything that matters for raw capability. It trails DeepSeek V4.1 Flash on DeepSWE by nearly thirty points. Reflection's own table shows this. They aren't hiding it. They're just not leading with it.

The pitch isn't raw capability. The pitch is efficiency. Beam matches GLM-5.2 on advanced reasoning benchmarks while using three to four times less inference compute, the company says. On SWE Bench Verified, Beam scores 80.9. That's a respectable number. It's not a frontier number. It's a workhorse number.

The Compute Math Is a Claim, Not a Receipt

Three to four times less inference compute. That's the sentence everyone repeats. Here's what it actually means.

GLM-5.2 activates 40 billion parameters out of roughly 744 billion total. Beam activates 23 billion out of 501 billion. The active parameter count is where the compute goes. Fewer active parameters means fewer floating-point operations per token generated. Fewer flops means less GPU time. Less GPU time means less money.

Reflection calls this "more intelligence per token." They say the model's forward-pass compute demands are lower than larger models like Qwen 3.8-Max, which requires significantly more computational resources per token.

All of this is true in theory. But here's the thing nobody's saying out loud: Beam has no published price. Every other model in that comparison table has a public list price you can plug into a spreadsheet. Beam doesn't. The platform console sits behind a signup wall. The blog post omits it. So any cost argument right now is an argument about compute, not about dollars. That's a real distinction.

The efficiency claim is an engineering claim, not a billing claim. It's like saying a car gets better gas mileage without telling anyone what gasoline costs where you live.

Training the Beast

They pretrained Beam in under four weeks on 6,144 Nvidia GB300 GPUs. Then the reinforcement learning run kicked in, four weeks on 10,500 GB300 GPUs. That run generated over 100 million rollouts. Reflection believes it's one of the largest such runs any open lab has attempted.

The RL phase did something strange. The model got better at browsing the web even though no browsing tasks were in the training mix. Given web access, it learned on its own to query other AI models and to use text-recognition tools to read documents. That's the kind of emergent behavior that makes researchers nervous and excited in equal measure. Nobody taught it that. It just started doing it.

They also trained a second model from the same base for safety and alignment, then merged the two. The safety results will appear in Beam's technical report. They plan to open-source the safety tests they built internally. Two government bodies are assessing the model, the US Center for Advancing Innovation and Standards for Super Intelligence, and the UK's AI Safety Institute.

The Money Behind the Machine

Reflection AI started in 2024. Two former Google DeepMind researchers, Misha Laskin and Ioannis Antonoglou, founded it in New York. Laskin worked on Gemini's training workflows. Antonoglou was a founding engineer on AlphaGo.

Since then they've pulled in approximately $4.7 billion in funding. Nvidia put in $800 million. Sequoia Capital, Lightspeed Venture Partners, and others followed. The most recent round put the pre-money valuation at $25 billion.

They signed compute deals this summer. A $6.3 billion deal with SpaceX for access to Nvidia GB300 chips at the Colossus 2 data center. A $1 billion deal with Nebius. They now have a pipeline of chips running through 2029.

The company is pitching Beam primarily at enterprises and sovereign governments. The argument is simple: if you're a government or a large company and you can't or won't use Chinese models, but you still want to own and control your AI, Beam is your option. Reflection calls these "AI factories", customized, locally deployed systems built on their models and trained on proprietary data.

Laskin told Semafor: "They don't really have very good options today".

Where Beam Sits in the Scramble

The open-weight landscape in October 2026 is crowded. Chinese labs lead. DeepSeek, Alibaba, Moonshot AI, Z.ai, they've been shipping open models that Western companies actually use. Amazon's AWS added Z.ai's GLM-5.3 to Bedrock. Moonshot's Kimi K3 landed on the platform last month.

The American response has been slow. Thinking Machines Lab shipped Inkling. Mistral keeps churning. Nvidia and Microsoft and Palantir and Meta signed a public letter in July calling for the US to support a domestic open-weight ecosystem. The Treasury Secretary urged American companies to build more open-weight models.

Beam is part of that push. It's not the best model in the room. It's the most efficient American model in the room. That's the pitch. That's always been the pitch.

Reflection is honest about the gap. Their own table shows Kimi K3 ahead on raw capability. They say Beam is "competitive" with GLM-5.2 and "approaching" Qwen 3.8-Max. Approaching. That word does a lot of work.

What Comes Next

The weights drop later this month. Apache 2.0 license. Technical report. Model card. Developer tooling. FP8 and NVFP4 quantized versions. Compatibility with widely used open-source libraries. Distribution through hyperscalers.

Beam is text-only. No image input. No audio. That's a difference from Inkling, which shipped multimodal from day one. If you need a model that reads images, Beam isn't it. Not yet.

They're already training the next model. Laskin says it will be "much more" powerful and performant than Beam. They're not standing still.

A Closing Observation

You watch a thing like this and you think about what it means. A company raises four and a half billion dollars. They burn through 10,500 GPUs for a month. They generate 100 million rollouts. They build a machine that activates a quarter of its brain per thought. And the whole thing hinges on a claim you can't verify until the weights come out and someone runs their own benchmarks.

That's the game. You make the claim. You ship the model. You let the market decide whether three-to-four-times-less-compute is a real number or a number that looks good on a slide.

Beam is out there now. A stranger in a bar at 2 a.m. It ordered something. It's talking. Whether anyone listens depends on what happens when the weights hit the street and the independent benchmarks start running.

The machine drifted in. It'll drift out or it'll stay. That part hasn't been written yet.