AI research wiki
ai-industry · 2026-10-18 · Tier 3

Interconnects: "Meta shakes up the frontiers, open and closed"

Interconnects: "Meta shakes up the frontiers, open and closed"

TL;DR. Nathan Lambert argues that the week of Muse Spark Open (Meta's 816B-total, 42B-active open model), Reflection's Beam and Mistral Large 4 (announced, weights not yet out) roughly halved the gap between the best American and Chinese open models. Without Muse Spark Open and Beam there would be 20+ Chinese open models better than the best Western one. He ranks the field Anthropic > OpenAI > Meta ≈ Gemini > xAI ≈ Chinese open weights > US open weights, and says the US frontier has narrowed to five labs.

The argument

  • Muse Spark Open should get adopted. Western open models get far more adoption than Chinese ones at the same performance-per-size, and demand for open inference is rising as open agents become real. Capability sits above GLM 5.2, near DeepSeek V4.1 Flash, behind GLM 5.3 and Kimi K3.
  • Probably a safety-tuned Muse Spark 1.2, not a broken 1.3. About a 1-point Artificial Analysis drop for the extra safety post-training.
  • The license is the risk. "Built with Muse" display, a revenue gate above $20M/month, an acceptable-use policy that also binds outputs and derivatives (so synthetic-data use is legally ambiguous). Lambert calls non-standard open licenses "an unnecessary strategic risk."
  • Organization beats resources. Meta's flat org (Alexandr Wang has 200+ direct reports) is shipping consistently. Labs fail from management layers; Llama failed from politics over performance.
  • The "tier 3" squeeze. Mistral, Nemotron, Reflection, Arcee struggle to move either frontier; frontier models now cost billions each. Microsoft could join the top five given its MAI scaling results.
  • Nobody can predict real-world model quality. Why did Claude go from Fable 5 (hit) to Opus 5 (dud) to Opus 5.5 (hit)? Evaluations are a proxy; the long tail is unknowable.
  • Valuation puzzle. Meta is valued under $2T, about where Anthropic's IPO is pitched; "one of the two valuations needs to correct massively."

Relation to prior wiki pages

  • Builds on the 10-16 Muse Spark Open coverage, which focused on its shipped serving stack (NVFP4 weights, DFlash2 drafter, MLX build).
  • DeepSeek's V4.1-Flash-1017 (10-18) landed the next day and moved the Chinese open frontier again, testing the "gap halved" claim within 24 hours.
  • The license argument meets AI Weekly's Tech Against Terrorism finding (10-17): stripped copies of open models now appear in under three days on average.

Source: Interconnects