Anthropic releases FlashProve with Lean formalization, proving 400B MoE inference on one H100
jiqizhixin · x · 2026-11-22
Anthropic released FlashProve, with an official technical paper and a Lean formalization of its inference-proof protocol. The protocol lets a model provider prove outputs of a committed model without disclosing weights: a hash-based protocol commits to weights and constructs a proof of greedy inference.
- A public-weight 400B integer MoE model was tested on a single H100, reporting 110 prefill tokens/s steady state; each output token costs 18.8 MB of proof.
- Soundness, completeness, and zero-knowledge are formally proved under stated assumptions and a fixed configuration.
- C++ and CUDA implementation benchmarks are explicitly separated from the formal theorems.
More from Models
- Tesla starts rolling out FSD v15 with 25x better safety-critical handling — ArnoCandel · 2026-11-22
- Joke tweet suggests Grok 4.5 vs 4.7 teams 'need to be studied' — zdhnarsil · 2026-11-22
- Simon Willison: Gemini 4 Hydrogen 'xhigh' thinking costs 6x more but is far better — film_girl · 2026-11-22
- Researcher argues LLM companies have no moat as Gemini 4 Neon kills the data flywheel thesis — adam_karvonen · 2026-11-22
- Gemini users say Low thinking still burns 5-hour quotas on basic prompts, want a non-thinking option back — patg84 · 2026-11-22
- GPT-6.2 Sweeps Top of Game Dev Arena; Astra Costs $0.98 Per Generation — gdb · 2026-11-22