← AI Switchboard
MODELS · October 17, 2026
Oct 16

Reflection releases the open weights of Beam, its 501-billion-parameter model, under Apache 2.0, as promised for October

The model card puts it at 81.6% on SWE-bench Verified in Reflection's own table. The FT reported on 10 October that Nvidia was in early talks to invest further in Reflection or buy it.

What changed: when Reflection announced Beam, its first open-weight model, on 5 October, the weights were promised for later in the month, and on 12 October they were still not out. On 16 October Reflection said: “Today, we have released Beam’s model weights on Hugging Face under an Apache 2.0 license.”

Reflection's Hugging Face page now holds Beam 1.0 repositories for the main model, FP8 and NVFP4 builds and a speculative-decoding drafter. The model card describes a sparse mixture-of-experts design with 501.3 billion total and 23.4 billion active parameters and a context of 1 million tokens. In Reflection's own table the full-precision build scores 81.6% on SWE-bench Verified, 80.1% on Terminal Bench v2.1 and 90.6% on GPQA Diamond; the table names no rival models. Reflection also released a safety report and Mirror CLI, the command-line harness it trained Beam with. The model card says a technical report is still to come, and the hosted API is in public beta with no published price.

Apache 2.0 is one of the most permissive licences: anyone can use, change and sell the model, including in commercial products.

The release comes a week after the Financial Times reported, according to Reuters, that Nvidia was in early talks to invest further in Reflection or acquire it, possibly through an arrangement that would bring over staff and license the technology. The FT reported, according to Reuters, that Nvidia has already invested $800m. Neither company responded to Reuters' request for comment then, and neither company's release material mentions the talks.

  • Confirmed Reflection's post says Beam's weights are out on Hugging Face under an Apache 2.0 licence. Reflection
  • Claimed The model card gives 501.3bn total and 23.4bn active parameters and, in Reflection's own table, 81.6% on SWE-bench Verified. Reflection model card, Hugging Face
  • Reported Nvidia is in early talks to invest further in Reflection or buy it, the Financial Times reported, according to Reuters; the FT reported that Nvidia has invested $800m. Reuters, via The Star

Models & releases

Today in the October 17, 2026 edition · front page