Reflection AI Beam: a 501B open-weight model aimed squarely at China's lead
Reflection AI, a Brooklyn startup founded in 2024 by two former Google DeepMind researchers, unveiled its first open-weight model, Beam, on October 5, 2026. Beam is a mixture-of-experts model with 501 billion total parameters, of which 23 billion are active per token, a 1-million-token context window and 23.8 trillion training tokens. Reflection says it matches Z.ai GLM-5.2 on advanced reasoning while using three to four times less inference compute, and beats leading Western open models. Weights are due under the Apache 2.0 license later in October; for now access is by waitlist. The performance claims have not been independently verified
The three lines
- Launch — October 5: 501B total / 23B active MoE, 1M-token context, 23.8T training tokens
- Claim — GLM-5.2-level reasoning at one-third to one-quarter the inference compute (unverified)
- Stakes — an Nvidia-backed U.S. lab takes on Chinese dominance of open models; Apache 2.0 weights due in October
Key questions
- Reflection AI Beam benchmarks
- **Company figures put it level with GLM-5.2 on advanced reasoning.** (not independently verified) | Item | Beam | GLM-5.2 (Z.ai) | |---|---|---| | Total parameters | 501B | ~744B | | Active parameters | 23B | 40B | | Context | 1M tokens | — | | Inference compute | 3–4x less (claimed) | Baseline |
- Who is Reflection AI
- **A 2024 Brooklyn startup founded by two ex-Google DeepMind researchers.** | Item | Detail | |---|---| | Funding | ~$4.7 billion total | | Valuation | $25 billion pre-money | | Backers | Nvidia, Sequoia, Lightspeed | | Compute | >$7B in deals with SpaceX and Nebius for GB300 chips through 2029 |
- When can I download Beam
- **Weights are scheduled for October 2026 under Apache 2.0.** | Stage | Timing | |---|---| | Announcement, waitlist | October 5 | | Red-teaming, final evals | Under way | | Weights and docs | Later in October (no date) | | Distribution | Hyperscalers and neoclouds |
For most of 2026, "best open model" has meant a Chinese model. An American lab is now challenging that directly. Nvidia-backed Reflection AI unveiled its first open-weight model, Beam, on October 5, 2026. It is large — 501 billion parameters — but the number Reflection stressed was compute: it says Beam matches a leading Chinese model while using a third to a quarter of the inference compute.
1. What Beam is
| Item | Detail |
|---|---|
| Announced | October 5, 2026 |
| Architecture | Mixture of experts (MoE), text only |
| Total parameters | 501 billion |
| Active parameters | 23 billion per token |
| Context window | 1 million tokens |
| Training data | 23.8 trillion tokens (public web + commercial) |
| Target work | Coding, reasoning, agents |
| License | Apache 2.0, weights due later in October |
The key ratio is 23 of 501. An MoE model holds many groups of "experts" but wakes only a few for each token, so it stores a huge model yet runs only about 5% of it per step. See "What mixture of experts is" for how that works.
2. The benchmark that matters: China
Reflection chose to measure itself not against U.S. models but against Z.ai's GLM-5.2 — a sign of how far Chinese labs have come to define open-model leadership.
| Item | Beam | GLM-5.2 |
|---|---|---|
| Total parameters | 501B | ~744B |
| Active parameters | 23B | 40B |
| Advanced reasoning | Parity (claimed) | Baseline |
| Inference compute | 3–4x less (claimed) | Baseline |
Fewer active parameters mean more requests per GPU, which matters as much as raw scores to companies that host open models themselves. Reflection also says Beam beat Thinking Machines Lab's Inkling on four coding benchmarks and leads Western open models on advanced reasoning. All of this is self-reported until the weights are out and others can test them.
3. Why Nvidia wants this to exist
Founded in 2024 by two former Google DeepMind researchers, Reflection has raised about $4.7 billion at a $25 billion pre-money valuation from Nvidia, Sequoia and Lightspeed, and signed more than $7 billion in compute deals with SpaceX and Nebius to secure Nvidia GB300 chips through 2029.
For Nvidia, a strong American open model is strategic. As Chinese open models increasingly optimise for domestic hardware, Nvidia benefits from a top-tier open model built to run best on its GPUs. Note what "open" means here: Beam publishes weights under a permissive license, not its full training data.
4. What remains unconfirmed
- Performance and efficiency claims await independent evaluation after the weights ship.
- The release date is only "later in October"; access is currently by waitlist.
- Non-English performance, including Korean, and the composition of training data are undisclosed.