Anthropic FlashProve: open zkSNARK prototype for provable LLM inference (~110 tok/s on Llama 4 Maverick H100)

Why it matters

【Key points】

• Anthropic (2026-11-02) open-sourced FlashProve: a zkSNARK prototype for provable inference—commit to weights, prove each reply came from that commitment without revealing weights
• CPU verifies.
• On Llama 4 Maverick (400B MoE, 4-bit, 1, 024-token): ~110 tok/s proving on one H100 (batch 24), ~19 MB proofs, ~70–95× GPU time vs vLLM; offline commit ~39 min.
• Ships report, C++/Lean verifiers, Lean completeness/soundness/ZK proofs.
• Limits: research prototype; fast code not formally tied to Lean; integer greedy MoE; architecture not hidden; 1,024 context.
• Related to but distinct from Anthropic’s earlier signed-traffic/recompute inference-verification prototype.

【Our read】

Background

• users, routers, evaluators, and regulators cannot tell from tokens which weights ran.
• FlashProve is a software-crypto path (vs TEE) Anthropic ties to pacing, tamper detection, provenance, and approved-model serving—an engineering step on public inference-integrity goals.

• Compare: open zkML was long unusable
• FlashProve’s Maverick numbers jump vs VerInf-class baselines, yet still ~two orders over native inference and MoE/short-context/integer-greedy bound.
• Complements e/1081 as serving-side attestation.
• TEE vs zkSNARK differ on trust roots; Anthropic flags the Lean↔executable gap.

Impact

• lifts zkLLM discourse and gives an open 400B MoE baseline; material for API trust, router audits, and future compliance toolkits.
• Watch longer context, dense models, cost cuts, and product defaults.
• Ship as paper—not a model release, and not “Claude is already fully provable in production.”

【Verifiable】

• PrimaryNov 2, 2026
• . Code
• X2026-11-03T20:10:08Z). Not on site (≤e/1083). Compare e/1081 (training EM) and e/1067 (OpenAI deployment monitoring