End-to-End Token Production with Domestic Chips
智东西 · wechat · 2026-07-17
This in-depth interview discusses how Qingcheng Jizhi uses a "front shop, back factory" approach to chain domestic chips, inference frameworks, and large models into a deployable Token production pipeline.
Core Thesis
- The bottleneck for domestic AI adoption isn't just chips and models; a missing inference engine is needed to convert compute into Tokens.
- Overseas inference engines are tailored for the Nvidia ecosystem. Migrating them directly to domestic chips causes issues with FP4/FP8 adaptation, VRAM usage, and precision loss.
What the Chitu Inference Engine Does
- Developed a software-based floating-point simulation architecture to preserve native model precision as much as possible.
- Optimizes mixed quantization, KV Cache scheduling, and multi-card parallelism based on domestic chip characteristics.
- Aims to enable domestic cards to produce Tokens more efficiently without requiring massive hardware overhauls.
Quantified Gains
- The article notes that deploying the full DeepSeek model using traditional open-source inference frameworks might require 4 domestic 8-card servers.
- Integrating Chitu reduces this to a single server, cutting overall hardware costs by roughly 75%.
AIPing and Bagua Furnace
- AIPing is a one-stop Token scheduling and service platform, unifying access to 30+ providers and 600+ model services.
- The Bagua Furnace (八卦炉) series targets Agent building, local private deployment, and compute center integration.
- The article emphasizes that their value lies in the flywheel of "evaluation—routing—call data," which continuously improves stability and success rates.
More from Infra
- Vercel AI Gateway data shows Anthropic, OpenAI and Google at 97.09% spend share — cramforce · 2026-07-21
- NVIDIA starts rolling out 102.4 Tbps Spectrum-6 switches for Vera Rubin AI factories — nvidia · 2026-07-21
- Mustafa Suleyman says Microsoft is preparing for an OpenAI exit, while a new chip costs 30% less than GB200 — thoefler · 2026-07-21
- Microsoft and Mistral sign multi-billion-dollar deal to expand AI infrastructure in Europe — The Decoder · 2026-07-21
- Speculative decoding boosts Qwen3.6-27B on one 5090, but slows crowded servers — luke_pacman · 2026-07-21
- NVIDIA says Blackwell Ultra hit 1,648 TFLOPs per GPU on DeepSeek-V3 671B training — NVIDIAAI · 2026-07-21