# Idemcore

> Verifiable AI inference. The same model and input give the identical result, bit for bit, on x86 and ARM CPUs, an Android phone and five NVIDIA GPU models across four generations (T4, A10, A10G, L40S, H100). Every answer can carry a signed receipt, and every decoding step is written to a documented hash chain, so anyone can re-run a result and check it.

This file is a plain-text summary of https://idemcore.net for AI assistants and readers. Every number below is copied from the site. Cross-device results measured 2026-10-30; H100 throughput benchmark measured 2026-11-01.

## Two ways to use it

1. Hosted API (public preview): OpenAI-compatible chat completions at https://api.idemcore.net, run on Idemcore's servers. Every answer carries a signed receipt anyone can verify. By default the API keeps hashes of each request and answer, not the text; storing full text is opt-in. Sign in at https://console.idemcore.net to create an API key. Docs: https://idemcore.net/docs
2. On-premises license: the engine runs on the customer's own servers, private cloud or devices; prompts and outputs stay there. Starts with a fixed-scope pilot and a written agreement.

## Key facts

- Bit-exact: every logit and every sampling step match across devices, not just the chosen token.
- 10/10 device pairs bit-identical against an x86-64 CPU reference (Qwen2.5-0.5B 8-bit, Qwen2.5-7B 8-bit and 4-bit, gpt-oss-20b, gpt-oss-120b).
- Single stream, Qwen2.5-7B 4-bit on an NVIDIA A10: 92.98 tok/s with receipts on, p50 10.77 ms, p99 10.87 ms (llama.cpp Q4_K_M 90.30 in a separate session). GPU step times.
- Batched on an A10G (Qwen2.5-7B 4-bit): 641 tok/s total at 8 streams, 1372 at 32, 1413 at 64 (GPU step times); every stream bit-identical to a single-stream x86 run. Measured 2026-10-30.
- Throughput benchmark, H100 80GB, Qwen2.5-7B (Idemcore 4-bit; vLLM and SGLang with AWQ), same client-side load generator for all engines, prefill included, aggregate tok/s at 32/64/128/256 concurrent requests. Measured 2026-11-01:
  - Idemcore: 3156 / 3988 / 4400 / 4637 (the build serving the API)
  - vLLM 0.32.0: 5708 / 8867 / 12838 / 12421
  - SGLang 0.5.23: 5048 / 6633 / 8676 / 9775
  - REE v0.8.0 (bf16, requests run one after another): 36 / 39 / 40 / 31
  - Throughput engines such as vLLM and SGLang are currently faster, about 1.6-2.9x Idemcore's aggregate throughput. Idemcore gives the same output regardless of batch (16 of 16 test prompts identical to their solo run at every batch size), with a verifiable receipt per answer. Numbers are updated as optimizations ship.
  - Method: byte-identical prompts from a deterministic scenario, exactly 256 greedy output tokens per request, N requests started together, 1 warm-up + 5 recorded rounds (median); official vLLM/SGLang containers, prefix caching off, otherwise defaults.
- Quality (512-position WikiText-2 slice): 7B 4-bit (g) perplexity +0.60%, top-1 agreement 95.5%; 8-bit -0.11%, 98.0%; gpt-oss-120b +6.23% (being worked on).

## Supported

- Models: Qwen2.5-0.5B, Qwen2.5-7B (8-bit and 4-bit), gpt-oss-20b, gpt-oss-120b. Other models on request, as part of a pilot.
- Hardware: x86-64 CPUs; ARM CPUs (Galaxy Z Fold5 / Snapdragon 8 Gen 2, Windows 11 ARM64 laptop / Snapdragon X Plus); NVIDIA T4, A10, A10G, L40S, H100. AMD and Apple GPUs are not supported yet.

## Pricing (USD, excluding taxes)

Paid plans open at launch. Nothing is billed during the public preview. Verifying a receipt is always free.

- Trial: $2 of credit for new accounts, no card
- Pay as you go: $0.15 per 1M input tokens, $0.60 per 1M output tokens
- Team: $99 per month
- On-premises license: $8K-15K per GPU per year
- Simulation & Replay: from $50K per year ($50K-400K); pilot $15K-50K
- Game & Esports Integrity: $25K-150K per title per year; pilot $15K-50K
- Banking & Finance: from $75K per year; pilot $25K-60K

Full list: https://idemcore.net/pricing. Refunds: https://idemcore.net/refund

## Products

- Cloud API (public preview): see "Two ways to use it" above. Free during the public preview, with preview limits.
- Free Verifier: re-run a logged inference and check it token by token. Free and open source (Apache-2.0). Repository: https://github.com/idemcore/idemcore-verifier. On PyPI: `pip install idemcore-verify` (https://pypi.org/project/idemcore-verify/)
- Python SDK: `pip install idemcore` (https://pypi.org/project/idemcore/), Python 3.9 or newer. API client with local receipt verification. Docs: https://idemcore.net/docs#sdk
- JavaScript/TypeScript SDK: `npm install @idemcore/sdk` (https://www.npmjs.com/package/@idemcore/sdk), Node 18 or newer, server-side. Same API and local receipt verification.
- Simulation & Replay (pilot): AI steps inside simulations replay identically, years later, on any machine. Details: https://idemcore.net/simulation
- Game & Esports Integrity (pilot): deterministic AI with a per-tick hash, server spot-check replay and evidence logs for match disputes. Or $0.0003-0.001 per MAU per month, whichever is greater; free for indie studios (under $1M annual revenue and under 100K MAU); tournaments $5K-30K per season. Details: https://idemcore.net/games
- Banking & Finance (pilot): replayable model decisions with signed receipts for audit and review, deployed on the bank's infrastructure. Details: https://idemcore.net/banking
- Coming soon (not available): Verifiable Compute Networks; Mobile / On-device SDK; Robotics, Autonomy & Aero; Chip Health & Energy.
- Every license starts with a pilot and a written agreement; final prices are set there. Licenses are not sold online.
- It does not replace an anti-cheat: it does not detect wallhacks or aimbots.

## Verify

Run the open verifier (https://github.com/idemcore/idemcore-verifier), Python 3.9 or newer:

    git clone https://github.com/idemcore/idemcore-verifier && cd idemcore-verifier && python3 -m idemcore_verify samples

Expected: "checked 65 logs (448904 token rows) and 58 pairs" and "PASS: every published hash recomputes and every pair is bit-identical". A match shows two runs agree; the per-step log hashes catch accidental differences, are not built to resist deliberate collisions, and do not prove which program produced a log. API receipts can be checked with POST /v1/receipts/verify or replayed: https://idemcore.net/verify

## Paper

"Idemcore: Bit-Identical LLM Inference from Phones to Datacenter GPUs, with an Open Per-Token Verification Protocol", Don Wick, 2026. Preprint v1.0 on Zenodo, DOI 10.5281/zenodo.23851129: https://doi.org/10.5281/zenodo.23851129

## Licensing

The verifier and log format are open source (Apache-2.0). The inference engine is closed and licensed per customer.

## Contact

hello@idemcore.net (pilots, questions, press); security@idemcore.net (security reports). https://idemcore.net/contact

## Pages

- https://idemcore.net/ (results, verify, plans, FAQ); Vietnamese: https://idemcore.net/vi/
- https://idemcore.net/pricing (prices)
- https://idemcore.net/simulation (Simulation & Replay)
- https://idemcore.net/games (Game & Esports Integrity)
- https://idemcore.net/banking (Banking & Finance)
- https://idemcore.net/docs (API documentation)
- https://idemcore.net/verify (check a receipt or a log)
- https://idemcore.net/status (service status)
- https://idemcore.net/security, https://idemcore.net/contact, https://idemcore.net/changelog
- https://idemcore.net/privacy, https://idemcore.net/terms, https://idemcore.net/refund, https://idemcore.net/cookies
- https://console.idemcore.net (sign in, API keys)
- https://idemcore.net/about
- https://idemcore.net/llms.txt (short index for AI assistants)