Reflection AI announces Beam (501B/23B active) open-weight reasoning model
C · Weak signal Aggregated news report of a vendor model announcement; benchmark and efficiency claims are the company's own and, as TechCrunch notes, not independently verified, with weights not yet released.
Summary based on the full text·Original: about 4 minutes to read
In short
A news summary from Let's Data Science (6 October 2026) of Reflection AI's 5 October launch of Beam, its first open-weight model, drawing on Reflection's announcement and on TechCrunch, Fortune, Business Insider and Sources.news. Beam is a text-only sparse mixture-of-experts model with 501 billion total and 23 billion active parameters, aimed at coding, reasoning and agent workloads. Reflection claims competitive benchmark results and three to four times lower inference compute than rival Western open models. Weights, model card and technical report are promised for later in October.
Key numbers
- 501B / 23Btotal and active parameters per inference in the sparse mixture-of-experts model
- 23.8 trillionpretraining tokens from web and licensed data
- 10,500Nvidia GB300 GPUs used for a four-week reinforcement learning run
- 80.9reported SWE Bench Verified score
As reported by the source.
English·Web page·opens letsdatascience.com
What it reports
- Architecture: sparse mixture-of-experts, 501B total parameters, 23B active per inference, text only.
- Training: 23.8 trillion pretraining tokens; reinforcement learning produced more than 100 million rollouts over four weeks on 10,500 Nvidia GB300 GPUs.
- Reported benchmarks: 77.2 on SWE Bench Pro v2-Hard, 80.1 on Terminal Bench v2.1, 80.9 on SWE Bench Verified.
- Reflection says Beam is competitive with Z.ai's GLM-5.2 on selected coding and agent benchmarks, while Kimi K3 remains ahead on raw capability.
- Claimed three to four times less inference compute than rival Western open models in its comparisons.
- TechCrunch reports a 1 million-token context window; Sources.news reports an intended Apache 2.0 licence.
- Status: final red-teaming and evaluation, early access by signup; weights and documentation due later in October 2026.
How they got it
The article compiles Reflection's launch post with three press reports. All performance figures are vendor-reported; no independent evaluation exists yet because the weights are not public.
Limits
- Benchmark and efficiency claims are unverified; scores depend on scaffolding, hardware, settings and possible contamination.
- No model card, safety evaluation or technical report was available at publication.
- The source is a secondary news aggregator, not the primary announcement.
Why it matters for companies
If the weights ship as described under Apache 2.0, companies gain another US-built open model strong enough for coding and agent tasks that can run on their own infrastructure. That matters for firms that cannot send data to external APIs or prefer not to rely on Chinese open models for policy reasons. A 501B-parameter model still needs substantial hardware to serve, so the cost case depends on measured deployment numbers once they exist. Evaluation on a company's own tasks should wait for the released weights and documentation.
Summary based on the full text. Read the original for the full argument and data.Read the original on letsdatascience.com
- Tags
- Model releases