Reflection Beam open weights are out: 501B model on Hugging Face under Apache 2.0, in BF16, FP8 and NVFP4

Reflection AI released the Reflection Beam open weights on October 16, 2026, making good on the promise it made when it announced the model on October 5. Beam is a 501-billion-parameter mixture-of-experts model with 23 billion parameters active per token, and the weights are on Hugging Face under the Apache 2.0 license. It joins a growing list of large open-weight models from US labs. When we first reported on Beam on October 12, the weights weren't out yet.

This story is based on Reflection's "Beam is now available" post, its Hugging Face collection and model card, Reflection's online model card and developer docs, OpenRouter's listing, and VentureBeat's October 16 report, checked October 18, 2026.

The short version

• Released: October 16, 2026, per Reflection's blog post.
• License: Apache 2.0 for the weights.
• Size: 501.3B total, 23.4B active parameters (model card). Text in, text out, mainly English. 1M-token context.
• Versions: BF16, FP8 and NVFP4, plus a 2B "DFlash2" drafter model for speculative decoding.
• Access: Reflection's own API (free public beta) and partners including AWS, Google Cloud, Microsoft Foundry, Oracle, Together AI, Fireworks, Ollama and OpenRouter.
• Nvidia: no deal announced. Reflection's CTO called the acquisition reports "baseless rumors."

Where can you download Reflection Beam?

From Reflection's Hugging Face collection, which lists four repositories:

Last verified: October 18, 2026

The model card lists the license as Apache 2.0. Reflection's online model card adds that it doesn't prohibit distillation, fine-tuning, quantization or merging. Apache 2.0 allows commercial use, modification and redistribution with notice requirements, and includes an explicit patent grant; our open-weight license comparison covers how that differs from the custom licenses on GLM and Kimi models.

Hugging Face shows 255B as the parameter count for the NVFP4 repository. That is a file-format artifact, not a smaller model: the repository's safetensors metadata counts most of the 4-bit weights as 246.6 billion 8-bit values, each holding two 4-bit numbers, which works out to the same 501B as the BF16 and FP8 repositories. The model card benchmarks all three formats as the same model.

What's in the model card?

The Hugging Face model card fills in details the October 5 announcement left out:

• Architecture: 52 layers (3 dense, 49 mixture-of-experts), 256 experts, gated grouped-query attention with sliding-window layers, and a 152K-token vocabulary.
• Context: 1M tokens.
• Knowledge cutoff: June 30, 2026.
• Chat format: a modified Harmony chat template.
• Safety: On Reflection's own "Iris" evaluations, Beam refused or redirected 88.5% to 89.7% of harmful requests and complied with 81.5% to 82.1% of harmless ones, depending on the format. A separate Beam Safety Technical Report is published on Reflection's site.

What hardware does Beam need?

Reflection hasn't published minimum requirements. Its developer docs walk through one setup "for development and evaluation": a single server with four Nvidia B300 GPUs running the NVFP4 weights plus the DFlash2 drafter in SGLang, vLLM or TokenSpeed, accepting up to 262,144 tokens of context (131,072 in TokenSpeed) and eight concurrent requests. That is a reference configuration, not a minimum. VentureBeat reported that the 8-bit and 4-bit versions fit on a single eight-GPU server with H100, H200 or GB300 chips, while the 16-bit version needs "tens of GPUs."

On Hugging Face, the repositories hold about 1.0 TB of files for BF16, 503 GB for FP8 and 294 GB for NVFP4. Only 23B parameters are active per token, but every expert has to be held in memory, plus the KV cache for long contexts.

How does Beam score now? (vendor-reported)

The model card reports Reflection's own results for each format, run at temperature 0.7 and top_p 0.99. It doesn't compare Beam with other models.

Last verified: October 18, 2026 (Reflection model card and "Introducing Beam" blog, updated October 8; vendor-reported)

The release numbers moved slightly in both directions from the October 8 tables we reported on October 12. The 4-bit version stays within 1.7 points of BF16 on these rows, which matters for anyone hoping to serve it on one machine. No independent evaluator had published Beam results as of October 18. For comparison, DeepSeek's own table puts its new V4.1-Flash-1017 at 77.0 on DeepSWE v1.1, though that's a different vendor's run, so the two shouldn't be read as one ranking.

Is Beam really more token-efficient than GLM-5.2?

That's Reflection's claim, and it's an estimate. Its October 5 blog says Beam "achieves scores comparable to GLM-5.2 while using 3-4x less inference compute" on advanced reasoning benchmarks. Reflection's footnote says it estimated compute as about 2 × active parameters × mean generated tokens per attempt, using token data from Artificial Analysis and DataCurve on DeepSWE, HLE and Terminal-Bench 2.1. It leaves out prefill, attention and serving overhead, and Reflection calls it "an approximate compute comparison rather than measured inference cost." The model card puts it differently: comparable solution quality to leading open models with about 3x fewer reasoning tokens. It doesn't name the models.

VentureBeat, reading Reflection's charts, said Beam used about 25,000 tokens per Terminal-Bench 2.1 attempt versus about 80,000 for GLM-5.2. Antonoglou told VentureBeat the training penalizes token use: "You don't give up anything. You just penalize it in terms of how many tokens it can use." Treat this as a rough cost proxy, not a measured bill.

What does Beam cost?

Reflection hasn't published prices. Its October 16 post says its own API at platform.reflection.ai is a public beta that's "free to start"; its developer docs describe per-minute and daily rate limits. OpenRouter lists "Beam 1.0 501B A23B," created October 16, with a 1.05M-token context and four hosts on October 18, each setting its own price per million input and output tokens:

For how these prices compare with frontier APIs, see our cost comparison.

Last verified: October 18, 2026

What happened with the Nvidia talks?

Nothing has been announced. On October 10, the Financial Times reported that Nvidia, already an investor, was in early talks to invest more in Reflection or acquire it, according to Reuters' account of the report. In VentureBeat's October 16 story, CTO Ioannis Antonoglou called the reports "baseless rumors," described Nvidia as a close partner and investor, and said: "We remain committed to our mission of building frontier open intelligence as an independent company." He didn't directly say whether talks are taking place. Nvidia hasn't commented publicly.

What's confirmed, reported and uncertain

• Confirmed (Reflection, Hugging Face, OpenRouter): the October 16 release, Apache 2.0 license, four repositories, model card specs and benchmark numbers, the free-to-start API beta, the partner list, the four-B300 reference setup, and the OpenRouter host prices.
• Reported: eight-GPU fit for 8-bit and 4-bit versions, chart readings of token use, and Antonoglou's comments (VentureBeat); Nvidia talks (FT, via Reuters).
• Vendor estimate: the "3-4x less inference compute" comparison with GLM-5.2.
• Uncertain: independent benchmark results, official hardware requirements, Reflection's eventual pricing, and the outcome of any Nvidia discussions.