FULL STORY

Reflection's Beam: From Leak to Launch to Backlash

After Axios leaked Nvidia-backed Reflection's plan to release its first open-weight model, the startup launched the 501B-parameter Beam on Oct 6. While some praised its efficiency, developers found it trailing DeepSeek and other Chinese models.

2026-10-04 ~ 2026-10-06 · 3 episodes · 45 posts

Episode 1 · Nvidia-backed Reflection to release first open-weight model, targeting top Chinese open-source rivals (2026-10-04, 14 posts)

According to an Axios exclusive, Nvidia-backed startup Reflection AI is set to release its first open-weight model, reportedly highly capable and aimed at challenging both US AI giants and emerging Chinese open-source rivals like DeepSeek and Qwen, though it is expected to initially trail the strongest US frontier closed-source systems. Notably, Axios also reports that a wave of Western open-weight models is coming, with several US labs expected to follow suit this month.

Confirmed

  • The source is an Axios exclusive citing informed sources: Reflection is preparing to release a powerful open-weight model to compete with top Chinese open-source models, though it will initially lag behind the strongest US frontier systems.
  • On compute, Reflection has been paying Elon Musk $150 million per month since July to rent Colossus capacity, and has committed over $7 billion in total compute investment.
  • Reflection is backed by Nvidia.
  • Axios reports more US labs will follow with open-weight model releases this month.

Not Yet Confirmed

  • Specific details such as the model's release date, parameter scale, and performance benchmarks have not been disclosed; Axios notes details are still emerging.
  • Claims about "threatening giants on both sides" are interpretive takes from media and reposters; actual competitiveness awaits validation once the model ships.

Why It Matters

  • If true, this marks another significant move by the US camp in open weights to benchmark against China's top open-source models, potentially reshaping the open-source landscape currently dominated by DeepSeek, Qwen, and others.
  • Reposter VraserX commented that the US market needs more serious open-weight models rather than more mid-sized 200B models, reflecting community demand for high-quality Western open-source models.
  • The $7+ billion compute commitment is a rare scale, showing how heavily the startup is betting, tied to its relationship with Nvidia.

Episode 2 · Reflection AI Unveils Beam, Its First Open-Weight 501B MoE Model (2026-10-06, 27 posts)

On October 6, Reflection AI (founded by Misha Laskin) officially announced Beam, its first open-weight model, with full weights set to be released later this month. Beam uses a sparse MoE architecture with 501B total parameters and only 23B active, positioning it as a "workhorse" open-source model for coding, agents, and scientific tasks.

Confirmed

  • Sparse MoE architecture with 501B total parameters and 23B active
  • Pre-trained from scratch on 23.8 trillion curated tokens (web data plus proprietary licensed data)
  • The team says it achieves high token efficiency at inference and advances the "Western open-source frontier" on coding and agentic tasks
  • At announcement it was in the final stage of red-teaming; weights will be released this month under the Apache 2.0 license
  • Target users are enterprises, governments, and developers

Why it matters

Beam is Reflection AI's first public model, entering the open-source ranks at the 500B-parameter scale and going head-to-head with today's leading open-source frontier models, offering a new option for enterprises and developers who need large-scale open models. Its Apache 2.0 licensing and this month's release timeline make its real-world availability worth watching.

7 more related posts →

Episode 3 · New Open-Weight Model Beam Falls Short of Chinese Rivals in Benchmarks (2026-10-06, 4 posts)

Startup Beam's first open-weight model claims to advance the Western open frontier and competes with GLM 5.2, but reviewers including Bindu Reddy report it lags behind Chinese models like DeepSeek Flash, Qwen, and GLM on key coding benchmarks.