Anthropic Launches Open-Source FlashProve for Model Identity at 110 Tokens/s
By Ali Al-Montaser — Senior AI Technical Editor
Anthropic has released FlashProve, an open-source system using zkSNARKs to cryptographically prove AI model identity without revealing weights. It achieves 110 tokens/second on a single H100 for Llama 4 Maverick (400B, 4-bit), with ~19 MB proofs per 1024 tokens—70-95x faster than running the model on vLLM. The Apache 2.0 release includes C++ and Lean code, protocol specs, and a technical report, enabling regulatory compliance and preventing secret model swaps.
Executive Overview
On November 3, 2026, Anthropic officially announced FlashProve, an open-source provable inference system that uses zkSNARK proofs to cryptographically verify that a response was generated by a specific AI model without revealing its weights. Released under the Apache 2.0 license, FlashProve achieves 110 tokens per second on a single NVIDIA H100 GPU for Llama 4 Maverick (400 billion parameters, 4-bit quantization), with a proof size of approximately 19 MB per 1024-token request—making it 70-95x faster than running the same model on vLLM. This breakthrough addresses a critical gap in AI transparency, enabling regulatory compliance and preventing secret model swaps in commercial deployments.
📊 Official Technical Specifications & Data Sheet
| Technical Dimension | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Free and open-source under Apache 2.0 (code, Lean proofs, documentation). No licensing fees or subscription plans. |
| 🌐 Platforms & Immediate Availability | GitHub (open-source code), technical report, protocol specification, interactive Lean theorems visualizer. Proof generation on H100 GPU; verification on local CPU. |
| ⚡ Performance & Speed Benchmarks | 110 tokens/second on a single H100 for Llama 4 Maverick (400B, 4-bit). Proof size ~19 MB per 1024 tokens. 70-95x faster than vLLM. |
| 🛡️ Security & Tamper Resistance | zkSNARK proofs do not reveal weights; cryptographic verification of model identity and resistance to model swapping or token flow manipulation. Lean proofs for formal verification of security guarantees. |
| 🧠 Context Window | Not applicable; the system is independent of model context size and generates a proof per 1024-token request. |
| 🌍 Arabic Language & Regional Support | Language-agnostic; verifies any text output including Arabic. No regional restrictions on open-source code. |
Deep-Dive Features & Architecture
FlashProve leverages zkSNARK (Zero-Knowledge Succinct Non-Interactive Argument of Knowledge) technology, which allows a model provider to commit to a specific model and then prove that every response originates from it without disclosing the weights. The system generates a proof of approximately 19 MB per 1024-token request, verifiable on a local device in seconds. The release includes five components: a blog post, a technical report, a protocol specification written by Claude, executable code (prover and verifier in C++ and Lean), and Lean proofs for higher-level mathematical statements with an interactive visualizer.
Crucially, the system preserves weight confidentiality through a cryptographic fingerprint that can be published publicly, making it applicable to closed commercial models like Claude and GPT. It also provides auditable Lean proofs—a formal verification language—to ensure that claimed security guarantees are mathematically sound. Researchers note that the system could prevent an AI company from secretly swapping in a cheaper model to cut costs or a more expensive model to inflate performance results.
The research team includes Roy Rinberg and Alex Allain from Anthropic; Alessandro Chiesa from EPFL and StarkWare; Miranda Christ from MIT and RESI; Rahul Dalal from the Anthropic STEM Fellows program and Succinct; Giacomo Fenzi from EPFL; Ben Mayer from the Anthropic Fellows program and MIT; and Guy Weissenberg from EPFL. Anthropic stated that the research is ongoing and there are currently no plans for commercial deployment.
Benchmark & Competitive Performance
According to the official announcement, previous open-source systems are orders of magnitude slower than FlashProve. On Llama 4 Maverick (400B, 4-bit quantization), FlashProve achieves 110 tokens/second on a single H100, while running the same model on vLLM requires 70-95x less time (i.e., proof generation consumes 70-95 times more GPU time compared to standard inference). This represents a significant advance in making inference proofs practical at scale.
Industry Impact & Enterprise Adoption
FlashProve offers a major opportunity for developers and enterprises to build trust layers around large language models used in government and financial applications. Organizations can verify that the model in use is indeed the approved model without exposing the provider's weights. Because the system is language-agnostic, token efficiency for Arabic text depends on the underlying model (e.g., Llama 4 Maverick) rather than FlashProve itself, meaning development costs for Arabic startups are limited to integration and verification infrastructure. The open-source nature and Apache 2.0 license lower barriers to adoption, enabling regulatory compliance and auditability in AI deployments.
Conclusion
Anthropic's FlashProve marks a pivotal step toward transparent and accountable AI. By combining zkSNARK cryptography with formal verification, it provides a practical, open-source tool for proving model identity at unprecedented speed. While commercial deployment is not yet planned, the release empowers developers worldwide—including in the Arabic-speaking region—to build trustworthy AI systems that resist secret model swaps and meet regulatory scrutiny.
Media Source: TechCrunch AI | Official Company Statement: Original Source | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
FlashProve is an open-source provable inference system based on zkSNARK proofs that lets users cryptographically verify that a response came from a specific AI model without revealing the model's weights. Anthropic released it on November 3, 2026, under the Apache 2.0 license.
On Llama 4 Maverick (400 billion parameters, 4-bit quantization), a single H100 GPU generates proofs at 110 tokens/second, with a proof size of about 19 MB per 1024-token request—70-95x faster than running the same model on vLLM.
FlashProve is open-source under Apache 2.0 on GitHub and does not expose model weights; it preserves weight confidentiality via a cryptographic fingerprint that can be published publicly. There are currently no plans for commercial deployment within Anthropic or any other lab, as it is medium-term research.
Proof generation requires at least one H100 GPU, while verification can be done on a local CPU in seconds. The release includes C++ and Lean executable code, protocol specifications, a technical report, and an interactive visualizer for Lean theorems.
FlashProve is not a language model but a language-agnostic cryptographic verification system, so it can verify any text output, including Arabic. There is no official data on specific Arabic text performance, but it works with multilingual models such as Llama 4 Maverick.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
