TrueOpen
An Open GPU Networkfor Verifiable and Private AI Inference
Cost-Efficient Verification Trustless Coordination High Concurrency
Our vision
Truly open AI also needs an open GPU network to run open-source models. TrueOpen connects independent GPU providers so people can use these models with independently verifiable computation and privacy protection for their inference data.
Verification
An answer alone cannot tell you whether a provider used the agreed model, precision, or computation. We are researching methods that let independent Verifiers check that work.
Verifiers independently check output-token log probabilities in parallel, without regenerating the answer, and reach majority consensus. In Qwen2.5-7B tests, three replays together used 1.9%–14.8% of generation GPU time. Similar efficiency gains are expected for larger dense models under comparable conditions, especially with long outputs; exact ratios vary.
SLP uses cryptographic proofs to verify inference for fixed-point quantized models. In TinyLlama tests, sampling reduced commitment and proof-generation time by approximately 78% versus full proving; separately, packing three requests reduced proving time by approximately 61% versus individual proofs. Similar savings are expected across model sizes under comparable model, proving, and workload configurations; exact reductions require measurement.
Privacy
A shielded pool and separate keys for each task reduce public links between users and their inference requests. Content encryption protects inputs and outputs.
Users spend from a shared shielded pool, proving spending authority without revealing which deposit funded the task. Service nodes can still confirm the budget is funded.
Users authorize tasks with a public-key address and a digital signature, never a name or home address. Each task uses its own keys, so requests are not grouped under one public user identifier.
Users can choose to encrypt inference content. Keys are delivered to authorized participants according to their task roles, including the Worker that needs the input to run inference.
Performance & Scaling
GPU nodes process multiple inference tasks in parallel, and Verifiers check compatible tasks in batches. The DA layer keeps task data available for verification, while batched blockchain submissions reduce transaction overhead.
Users receive results before verification and settlement finish. Verifiers group tasks by Worker, model, and verification rules for batch recomputation.
The DA layer stores inputs, outputs, and computation evidence for verification and dispute resolution. Task records are submitted to the blockchain in batches to reduce transaction overhead, while each record is checked individually.
TrueOpen research
With 1,000 output tokens, the ratio rose from 1.9% to 14.8% as input length increased from 50 to 8,000 tokens. Measured on one RTX PRO 6000 Blackwell GPU at batch size 1; excludes network and settlement overhead.
Read the experiment summary Verification cost / Qwen3-8B & 32BOne verifier prefill cost about 1% of generation time at 128 input tokens and about 8% at 8,192 — the same band on Qwen3-8B (RTX 4090) and on Qwen3-32B (H100 NVL), 96 combinations in all. The worker pays for every token in serial decode; the verifier reads every position in one parallel pass. For single-token outputs the verifier is the slower of the two.
Read the experiment Dense models / Qwen3-32B & 8BIn the Qwen3-32B tests, logprob checks distinguished the original BF16 model from quantized and smaller-model substitutes across the tested GPU hardware. Verifiers replayed the existing output rather than generating it again. The report does not quantify verification GPU time.
Read the experiment MoE models / Qwen3.6-35B-A3BIn the Qwen3.6-35B-A3B tests, logprobs alone were not a robust identity check. Layer-0 expert routing separated BF16 from FP8 more clearly than routing across all layers. The report does not quantify verification GPU time.
Read the experiment Cryptographic proofs / GPT-2, TinyLlama-1.1B, Llama-2-70BSealing 163 chunk boundaries and proving 5 of them took 21 minutes on a 2 TB CPU host; the proof is 4.34 MiB. All seven adversarial cases were rejected. Single-audit detection equals sampling coverage, which is 1.9% at this setting.
Read the experimentThe 7B benchmark measures computation time using three verification replays, while the Qwen3-8B and Qwen3-32B benchmark measures one; the Dense and MoE experiments study model identity; the SLP report covers cryptographic proofs. Results apply to their reported test conditions.
View all research and referencesHow it works
Development
The sampled layerwise proof of concept is complete. The next stage focuses on security hardening and a testnet prototype, followed by performance improvements and production deployment.
View the development roadmapJoin us
We are a practical, straightforward team. If you have deep expertise in cryptography, blockchain, or AI, we welcome your contribution.