- On October 5, 2026, Reflection AI announced Beam, a text-only mixture-of-experts model with 501B total and 23B active parameters. It says the weights will ship under Apache 2.0 "later this month"; as of October 12 they aren't on Hugging Face.
- Reflection's own benchmark tables (updated October 8) put Beam behind Kimi K3, GLM 5.3 and DeepSeek-V4.1-Flash on most shared coding and terminal tests. Its pitch is efficiency: "3–4× less inference compute" than GLM-5.2 for similar reasoning scores, a vendor estimate.
- On October 10, the Financial Times reported Nvidia is in early talks to invest further in Reflection or acquire it, possibly as an acqui-hire. Reuters couldn't verify the report, and neither company has commented.
Contents
On October 5, 2026, Reflection AI announced Beam, which it calls its first open-weight model: a 501-billion-parameter mixture-of-experts model with 23 billion parameters active per token, built for coding and agent work. Beam has been announced, not released. Reflection says the weights, technical report and model card will come "later this month" under an Apache 2.0 license, and until then access is through a waitlist. Five days after the announcement, on October 10, the Financial Times reported that Nvidia, already an investor, is in early talks to put more money into Reflection or to buy it.
This story is based on Reflection's announcement, TechCrunch's launch coverage, and Reuters' and SiliconANGLE's accounts of the FT report, checked October 12, 2026.
The short version
- What: Beam, a text-only sparse MoE model. 501B total parameters, 23B active, and a 1M-token context according to Reflection.
- Status: Early access by waitlist. Reflection says the model is "undergoing final red-teaming and evaluations".
- License: Apache 2.0, as promised by Reflection. Nothing has been published under it yet.
- Benchmarks: All vendor-reported. On most shared tests Beam trails the largest Chinese open models, and Reflection's pitch is compute efficiency.
- Nvidia: The FT reports early talks on a bigger stake, a purchase or an acqui-hire. Unconfirmed.
What is Reflection Beam?
Reflection's announcement describes Beam as a sparse mixture-of-experts model with 52 layers and interleaved local and global attention. The company says it trained Beam from scratch.
- Pretraining: 23.8 trillion tokens on 6,144 Nvidia GB300 GPUs, in "under four weeks."
- Reinforcement learning: about 10.5K GB300 GPUs for four weeks, more than 100 million rollouts, and "nearly one million" environments.
- Context: RL ran at up to 256K tokens. Midtraining "extends Beam's effective context length to 1M tokens," according to Reflection.
- Modality: Text only. Reflection says Beam "is not multimodal."
Reflection pitches Beam as a "workhorse model" for enterprises, the public sector and developers, against cheaper Chinese open models such as DeepSeek-V4.1-Flash and Kimi K3.
When can you download Beam, and under what license?
Not yet. The announcement says: "We will release the weights, technical report, model card, and developer artifacts later this month," and that the weights will come "under an Apache 2.0 license." As of October 12, 2026, Reflection's blog had no follow-up post and we found no Beam repository from Reflection on Hugging Face. For now, a waitlist at platform.reflection.ai gives "a select group of users" early access. Reflection hasn't published API pricing.
If it ships as promised, Apache 2.0 is among the most permissive terms a large model has carried: commercial use, modification and redistribution with notice requirements and an explicit patent grant. Our open-weight licenses comparison covers how that differs from the custom licenses on Kimi and GLM models.
What hardware does Beam need?
Reflection hasn't published inference requirements. That's expected to come with the model card. As an illustration only, using our arithmetic rather than a Reflection figure: storing 501 billion parameters takes about 1 TB at 16-bit precision and about 500 GB at 8-bit, before the KV cache and serving overhead. Only 23B parameters are active per token, but memory is set by the total, so Beam is a multi-GPU server model, not a laptop one. Our cost comparison of frontier APIs and open models explains why hosted access often beats self-hosting at this size.
How does Beam score? (vendor-reported)
Every number below comes from Reflection's own tables, last updated October 8, 2026. Reflection doesn't fully describe its harnesses or settings, and no independent lab had published Beam results as of October 12. The table shows only shared rows where Reflection lists a score for every model shown.
| Benchmark (Reflection's label) | Beam | GLM 5.3 | Kimi K3 | DeepSeek V4.1 Flash |
|---|---|---|---|---|
| Terminal Bench v2.1 | 80.1 | 88.2 | 88.3 | 90.6 |
| DeepSWE v1.1 | 44.4 | 61.0 | 68.0 | 74.2 |
Last verified: October 12, 2026 (Reflection, "Introducing Beam"; vendor-reported)
On the rows Reflection reports for both, Beam beats Thinking Machines Lab's Inkling: 80.9 vs 77.6 on SWE-bench Verified and 65.5 vs 54.3 on SWE-Bench Pro v1. TechCrunch notes that Inkling is multimodal and Beam is not. Reflection is candid about the gap at the top: "Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time."
That efficiency claim is the core of the pitch. Reflection says Beam matches GLM-5.2 on advanced reasoning benchmarks with "3–4× less inference compute." The figure is Reflection's estimate, computed as roughly 2 × active parameters × mean generated tokens. The company notes that this leaves out prefill, attention and serving overhead. Treat it as a rough proxy for cost, not a measured bill. Benchmark versions also matter here: Reflection reports Terminal Bench v2.1, while Terminal-Bench itself is now on version 4.0. Our guide to coding-agent evaluations explains why scores from different harness versions don't line up.
What did the FT report about Nvidia and Reflection?
According to Reuters' and SiliconANGLE's summaries of the paywalled FT story, published Saturday, October 10, 2026:
- Nvidia is in early talks either to deepen its investment in Reflection or to acquire it.
- One possible structure is an acqui-hire: Nvidia would hire Reflection's staff and license its technology without buying the company. The FT said this could avoid a lengthy regulatory review.
- An agreement could come "within weeks," but the talks could still collapse.
- Nvidia has invested $800 million in Reflection so far, according to the FT. SiliconANGLE says Nvidia led Reflection's $2 billion round in October 2025.
Reuters said it "could not immediately verify the report," and neither company had commented as of October 12. CEO Misha Laskin told CNBC in April that Reflection's latest round closed at a $25 billion pre-money valuation. TechCrunch, citing PitchBook, puts total funding at about $4.7 billion, with Sequoia and Lightspeed among the other backers.
Nvidia also supplies the hardware: Beam was trained on GB300 systems Reflection rents from SpaceX and Nebius under deals TechCrunch values at more than $7 billion through 2029. Rival chipmakers are buying labs too. On September 28, AMD agreed to buy World Labs for about $8.2 billion in stock.
What's confirmed, reported and uncertain
- Confirmed (Reflection): Beam's architecture and size (501B/23B), training figures, text-only modality, waitlist access, the Apache 2.0 promise, the "later this month" timing, and the benchmark tables as published October 8.
- Reported (FT via Reuters and SiliconANGLE): Nvidia's early talks and the possible deal structures, the $800 million invested, and a possible agreement within weeks.
- Reported (TechCrunch, CNBC, PitchBook): funding totals, the $25 billion pre-money valuation, and the compute deals.
- Uncertain: whether and when the weights ship, inference hardware needs and pricing, independent benchmark results, and whether any Nvidia deal happens.
What it means for you
If you want a permissively licensed, US-built model for coding agents, Beam is worth watching, but there's nothing to download yet. Wait for the model card and independent results. Apache 2.0 weights can't be withdrawn once released, but whoever controls Reflection will decide the terms of future versions.
About this storyBased on the sources linked below. Editorial standards
