| 266402X | High-performance detection technology | Adds to existing 7 items: deepfake video/audio/image/multimodal detection systems ≥98% accuracy & recall; steganography/extraction with BER ≤0.5% or recall ≥99.9%, FPR ≤0.1%, SSIM ≥0.999 / PSNR ≥41 dB / LPIPS ≤0.005 |
| 266404X | Information-confrontation technology | Adds AI-assisted vulnerability discovery: benchmark exploitability scoring ≥75%; end-to-end exploit-chain success ≥50% (mainstream OS/browser) or ≥40% (arbitrary code execution); ≥10× senior-expert efficiency or 10× more exploitable-vulnerability findings; autonomously chains ≥2 low/medium flaws into a runnable multi-stage PoC; autonomously finds ≥5 previously unknown high-severity flaws (or ≥50 unknown flaws) in one major open-source project or mainstream OS in a single run; top tier of national/international exploit-capability grading |
| 266501X | Information-processing technology | Adds to existing 18 items: (19) voiceprint recognition — TPR ≥99.1% at FPR ≤0.01%, latency ≤300 ms with throughput ≥100 req/s, spoof-detection ≥98% or FAR ≤1%, plus minority-language/dialect optimisation ≥90%; (20) translation for Chinese and minority languages — human MOS ≥4.8, semantic fidelity ≥95% with BLEU ≥35 (general) or ≥30 (professional domains) |
| 266504X | AI large-model technology | (1) Large models trained on China's "important data" (重要数据) and then serving; (2) core large-model technology — weights, key checkpoints, underlying source code, architecture design, training optimisation, inference acceleration — meeting any of: compute ≥10²⁶ FLOPs (rendered in extracted text as 1026次浮点运算), parameters ≥20,000亿 (= 2 trillion), or "globally leading capability" — defined as global top-10 at export time in public evaluations by mainstream bodies/platforms across comprehensive performance, specialised capability, global call volume, safety and training efficiency. Exclusion: models trained with only a ≤0.1% share of important data, de-identified, not reproducible in output, or lawfully public |
| 266505X | AI agent technology | (1) Autonomous planning & task decomposition: natural-language instruction split into ≥20-step executable subtask chains, decomposition accuracy ≥80%, average subtask success ≥90%; (2) multi-agent collaboration architecture: ≥60 concurrent agents, total completion time ≤2× the fastest single-agent time, answer agreement ≥85%; (3) tool use & system operation: ≥200 integrated third-party tools, tool-selection accuracy ≥90%, end-to-end success ≥90% across 3+ tool-type combinations; (4) AI game/decision technology: decision depth ≥20 steps or ≥10 concurrent game agents, adaptive strategy with ≤200 ms latency and ≥10 control units, exceeding 90% of humans on standard game benchmarks |
| 266506X | AI chip-design technology | (1) AI processing-chip design (architecture, instruction sets, IP cores, software stacks) meeting any: ≤14 nm process, floating-point/integer compute, 2.5D/3D advanced packaging, high-speed interconnect, non-Von-Neumann architectures (compute-in-memory, neuromorphic), or use in autonomous driving/CV/speech; (2) memory-chip design — 3D NAND: density ≥8 Gb/mm², I/O ≥1.5 Gbps, or bonding; DRAM: ≤18 nm process, vertical-channel transistor, bonding, or multilayer stacking |
| 266507X | AI compute-cluster technology | (1) chip/node interconnect bidirectional bandwidth ≥800 GB/s; (2) distributed-training fault tolerance: MTTR ≤120 s, visible interruption ≤30 s, GPU effective-compute waste ≤5%; (3) communication scheduling: linear scaling ≥98%, end-to-end latency ≤1 µs, comm-hiding ≥90%; (4) heterogeneous scheduling across ≥2 vendors' AI chips: bandwidth utilisation ≥90% or cross-node latency ≤5 µs, throughput +≥30% / utilisation +≥65% / cost −≥20% / MTTR ≤10 min with availability ≥99.9%, cross-geography scheduling efficiency ≥80%. Note: "large-scale intelligent compute cluster" = >10,000 AI accelerator cards |
