FlashProve
Anthropic's open-source zkSNARK prototype proving a hosted model's output came from committed weights without revealing them.
This entry is marked for review. Check the original sources before relying on its details.
Overview
Anthropic released FlashProve on November 2, 2026 as an open-source research prototype that attaches a zero-knowledge proof to each batch of model output, so an external verifier can check it came from committed weights. It is software-only and needs no trusted hardware. Anthropic demonstrated it on Llama 4 Maverick (400B MoE, 4-bit) at about 110 tokens per second of proof generation, and labels it a research artifact, not production-ready.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- zkSNARK proof per batch tying output to a weight commitment
- Software-only design, no trusted hardware
- Demonstrated on Llama 4 Maverick (400B MoE, 4-bit) at about 110 tokens/second proving
- Verification on a CPU after downloading a roughly 6.5 MB commitment
- Release reported to include a technical report, protocol spec, prover, two verifiers and Lean proofs; Apache 2.0
- Proving costs 70-95x more GPU time than inference and supports 1,024-token contexts
Use cases
- Checking that an API provider serves the advertised model
- Independent evaluations of hosted models
- Research on regulatory approved-model enforcement
Pricing
Free, open source (reported Apache 2.0); research prototype with no commercial offering.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 2 Nov 2026
- Record updated
- 7 Nov 2026
- Record last checked
- 7 Nov 2026