Reflection's Beam: a real open-weight rival, or a scorecard until the weights ship?

On October 5, New York lab Reflection AI announced Beam, its first model: a 501-billion-parameter open-weight system pitched as a Western answer to the Chinese open models that dominate self-hosted AI. The catch is that nobody outside a waitlist can run it yet. So the question for anyone choosing an open model for coding agents, internal tools or sovereign deployments is simple: is Beam a real alternative today, or just a scorecard until the weights ship?

The short answer

Beam is a credible entry on paper, and its most interesting claim is about efficiency rather than raw capability. But every number so far is vendor-reported, the weights are promised for "later this month," and on Reflection's own charts Beam trails the newest Chinese open models on most agentic coding tests. Treat it as a strong candidate for your evaluation list, not a migration target, until the weights and technical report are public and independent tests land.

What Reflection actually announced

According to TechCrunch and DataCamp's breakdown of the announcement, Beam is a text-only sparse mixture-of-experts model with 501 billion total parameters, of which 23 billion are active for each token. Reflection says it was pretrained on 23.8 trillion tokens and supports a 1-million-token context window.

The training disclosures are unusually detailed for a launch post. Reported figures include:

• Pretraining: under four weeks on 6,144 Nvidia GB300 GPUs, per Implicator and DataCamp.
• Reinforcement learning: a further four weeks on roughly 10,500 GB300 GPUs, producing more than 100 million rollouts across about a million coding, agentic and STEM tasks.
• License: Apache 2.0 is planned, with weights, a technical report and developer tooling due later in October.

Access today is a private beta. DataCamp reports an OpenAI Chat Completions-compatible API under the model ID Beam-501B-A23B, a beta cap of 262,144 tokens of combined input and output (well below the advertised 1 million), and no published pricing as of October 6.

Where Beam leads and where it trails

Reflection's comparison tables, as reproduced by DataCamp, MarkTechPost and Implicator, tell a more nuanced story than the headline framing.

Against Western open models

Beam's clearest win is over other US open-weight systems. On SWE-bench Verified it reports 80.9, against 70.7 for Nvidia's Nemotron 3 Ultra. That matters for organisations that are barred, by policy or procurement rules, from deploying Chinese-origin weights.

Against Chinese open models

Here the picture is mixed. Beam is roughly level with Z.ai's GLM-5.2 on several tests: 80.1 versus 81.0 on Terminal-Bench 2.1, 44.4 versus 44.0 on DeepSWE v1.1, and 65.5 versus 62.1 on SWE-bench Pro v1. But it trails newer models clearly. On Terminal-Bench 2.1, GLM-5.3 scores 88.2, Moonshot's Kimi K3 88.3 and DeepSeek V4.1 Flash 90.6. On Humanity's Last Exam without tools, Beam's 36.2 sits below GLM-5.2's 40.5 and Kimi K3's 46.9. On BrowseComp, Beam reports 77.4 against Kimi K3's 91.2.

In other words, Beam is benchmarked closest to a model Semafor says was released in June, while the current Chinese frontier has already moved on. Implicator notes that Reflection itself acknowledges Kimi K3 remains ahead on raw capability.

The efficiency claim, examined

Reflection's central pitch is that Beam matches GLM-5.2 on reasoning "while using 3 to 4x less inference compute." This is plausible in direction: Beam activates 23 billion parameters per token, while MarkTechPost's comparison table lists GLM-5.2 at roughly 40 billion active. Fewer active parameters means fewer floating-point operations per generated token.

But three caveats matter. First, Reflection describes the figure as an approximate compute comparison, not a measured inference cost. Second, per RuntimeWire, the estimate excludes prompt prefill, context-dependent attention and serving overhead, which dominate costs in long-context agent workloads. Third, DataCamp's author points out that a FLOPs-per-attempt framing naturally favours sparse models like Beam.

There is also a hardware reality the efficiency framing does not erase. Active parameters set compute per token; total parameters set memory. By our own arithmetic, 501 billion parameters at one byte each (FP8) is roughly 500 GB of weights before any cache, and about half that at 4-bit precision. Reflection has said FP8 and NVFP4 variants are planned, according to MarkTechPost. Either way, this is multi-GPU server territory, not a workstation model.

Why it matters beyond benchmarks

The strategic case is about supply, not scores. CEO Misha Laskin told Semafor that businesses and governments wanting capable open models without Chinese origins "don't really have very good options today." TechCrunch reports Reflection has raised roughly $4.7 billion, at a $25 billion pre-money valuation, with Nvidia among its backers, and has signed more than $7 billion in compute agreements. A well-funded US lab committing to Apache 2.0 weights at this scale is meaningful for buyers who need a permissive license and a domestic supply chain, even if Beam is not the strongest open model overall.

For practitioners, the useful metric is the one DataCamp suggests: cost per solved task on your own workload. A model that scores a few points lower but runs at a fraction of the compute can win in production, but only if its serving overhead behaves as advertised.

What we know vs. what we're inferring

Known (reported by multiple outlets citing Reflection): Beam's 501B total and 23B active parameters; 23.8 trillion pretraining tokens; the GB300 GPU counts; the planned Apache 2.0 license; weights promised for later in October; access currently limited to a waitlist; the benchmark figures quoted above.

Not yet verified: every benchmark and the 3-4x compute claim. TechCrunch states plainly that the performance claims have not been independently verified.

Our inference: that memory requirements will limit Beam to multi-GPU servers; that its real-world cost advantage will shrink on long-context workloads because of the excluded prefill and attention costs; and that its strongest market is organisations that rule out Chinese-origin weights.

What to watch next

• By October 31, 2026 (11:59 p.m. ET / 12:59 a.m. UTC-3 on November 1): whether Reflection meets its "later this month" promise for weights, the technical report and the model card under Apache 2.0.
• Independent evaluations: once weights are public, third-party results on Terminal-Bench and SWE-bench variants, and measured throughput per GPU, will test the efficiency claim.
• Pricing: whether the API publishes per-token rates, and whether hosting partners undercut them.
• Context window: whether the 262,144-token beta cap rises toward the advertised 1 million.

Educational information only.