Most open-weight (开源权重) coding models tell you what they beat on a benchmark. Zhipu's GLM-4.6 tells you where it can run. The answer — on Chinese-made AI chips, at a fraction of the usual cost — is the more interesting part.
A coding model that learned to use tools mid-thought
GLM-4.6 was released and open-sourced on September 30, 2025, as the latest flagship in Zhipu's GLM series. Its pitch is narrow and practical: be a genuinely good coding and agent (智能体) model that enterprises can actually host.
The concrete upgrades over its predecessor:
- Context window expanded from 128K to 200K tokens, enough for long codebases and multi-step agent runs.
- Coding performance near Claude Sonnet 4; in Zhipu's own CC-Bench real-world coding tests it reached roughly a 48.6% win rate against Sonnet 4, near parity and ahead of other open baselines.
- Tool use during reasoning — the model can call search, run code, and read files inside its thinking trace, not only after it finishes a chain of thought.
- About 15% fewer tokens needed to finish the same task versus GLM-4.5, by Zhipu's measurement — capability and efficiency moving together.
Zhipu reports the model as a sparse mixture-of-experts (混合专家, MoE) design of roughly 355 billion total parameters with 32 billion activated, under an MIT licence, available on Hugging Face, ModelScope, and Zhipu's own MaaS platform.
The domestic-chip angle
The detail that separates GLM-4.6 from a routine model refresh is where it runs. Zhipu confirmed the model was adapted to Cambricon (寒武纪) chips using an FP8 + Int4 hybrid quantization deployment — described as the first FP8+Int4 model-chip integrated solution running on domestic silicon. It was also adapted to Moore Threads (摩尔线程) GPUs through the vLLM framework at native FP8 precision.
That matters because inference cost, not training cost, is what decides whether a small company can ship AI. Running a strong open model on locally sourced hardware removes two dependencies at once: a foreign cloud bill and a foreign chip supply chain.
What it costs to use
Zhipu bundled GLM-4.6 with an upgraded GLM Coding Plan: the company says it delivers Claude-level coding at roughly 1/7 the price with 3× the usage quota. The entry plan started around ¥20 per month — about US$2.8 / HK$22.
On public leaderboards at release, GLM-4.6 placed around fourth overall on LMArena and tied for China's top spot with Alibaba's Qwen3-Max-Preview, while beating DeepSeek-V3.2-Exp on coding. Zhipu is candid that it still lags Claude Sonnet 4.5 on the hardest coding tasks.
Where GLM-4.6 sits in the open-weight coding race
Zhipu is not alone in chasing this lane. Alibaba's Qwen, DeepSeek, and Moonshot have all shipped open-weight models aimed at coding and agents (智能体), and the competition has pushed context windows and tool-use quality upward across the board. GLM-4.6's distinct move is pairing that capability with a domestic-chip deployment story, which matters most for buyers who cannot or will not rely on US cloud infrastructure.
For a typical Chinese software team, the practical question is no longer "can an open model code?" but "can we host it cheaply and compliantly?" — and GLM-4.6 is built to answer yes. The model-card transparency (MIT licence, published weights, CC-Bench trajectory data released for community study) reinforces that positioning: it is meant to be inspected, fine-tuned, and run on your own metal.
Beyond the headline
GLM-4.6's domestic-chip story is best understood against the older assumption that a serious model had to run on Nvidia (英伟达) hardware. Once that assumption loosens, a whole class of buyers — government agencies, state-linked enterprises, and smaller firms wary of export controls — gains a credible path to sovereign inference. The catch is that "runs on domestic chips" and "runs as well as on Nvidia" are different claims; quantization tricks that shrink memory can also shave quality on the hardest tasks, and stability at scale is still being proven.
For global readers, the relevance is strategic rather than immediate. Most Western teams will not switch silicon to buy this model; what they should watch is the precedent. If a strong open model can be decoupled from a single supplier's GPUs, the moat around high-end accelerators looks less absolute — and that changes how everyone prices inference.
Trend reading
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the push to run strong open-weight models on domestic chips is less about patriotism than about cutting the cost floor that decides whether smaller firms can actually deploy AI.
Honest limitations
- GLM-4.6 shipped in September 2025; Zhipu has since released newer flagships (GLM-5 in early 2026), so this is not the current frontier from the same lab.
- The "15% fewer tokens" and "1/7 the price" claims are Zhipu's own; independent replication varies by workload.
- Coding ability, while strong for an open model, still trails the best closed systems on complex software-engineering benchmarks.
- Domestic-chip deployment figures are vendor-reported; real-world throughput and stability at scale are not yet broadly published by neutral third parties.
What readers can do now
- Test the coding plan before self-hosting. Start with Zhipu's GLM Coding Plan (≈ ¥20/month, about US$2.8 / HK$22) inside Claude Code, Cline, or Kilo Code to see if the coding quality fits your stack.
- If you control infrastructure, try the domestic-chip build. The Cambricon FP8+Int4 and Moore Threads vLLM paths are the differentiator — evaluate them if supply-chain or cost constraints matter to you.
- Measure token efficiency on real tasks. Run a fixed coding or agent workflow through GLM-4.6 and a closed alternative, then compare completed-task-per-dollar rather than headline scores.