跪拜 Guibai
← All articles
ChatGLM (Zhipu AI) · AIGC · Artificial Intelligence

GLM-5.3-Flash Matches Claude Opus 4.8 at 1/40th the Cost, Runs on Chinese Silicon

By 王若风 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A frontier model matching Opus 4.8 at 1/40th the cost, running on non-NVIDIA silicon under an MIT license, resets the price anchor for high-intelligence API access. Developers who previously rationed flagship-model calls can now run agents overnight without a rent-sized bill, and the open weights mean self-hosting is a real option.

Summary

GLM-5.3-Flash, the first natively multimodal model in Zhipu AI's GLM-5 series, hits a 57 on the Artificial Analysis Composite Intelligence Index — dead even with Anthropic's Claude Opus 4.8 — at a promotional price of 4.5 cents per task, roughly 1/40th the cost. The 320B-parameter model activates only 18B parameters across 45 layers, using a hybrid sparse-and-linear attention architecture that structurally drives down long-context serving costs. It ships with a 1M-token context window, an MIT license, and immediate availability across HuggingFace, API endpoints, and Zhipu's own coding and chat platforms.

Before the public launch, the model ran anonymously as 'Ox-Alpha' on OpenCode and OpenRouter, where it became the most-called model on both platforms that week — all traffic served entirely by a domestic Chinese chip cluster. Zhipu built a custom inference engine on SGLang with EPD disaggregation, W8A8 quantization, and mixed-precision caching, achieving per-token costs and hardware efficiency comparable to mainstream NVIDIA GPU deployments.

On coding and agent benchmarks, GLM-5.3-Flash surpasses its larger predecessor GLM-5.2 across the board and edges past Opus 4.8 on DeepSWE, Toolathlon, and AutomationBench. Native multimodality lets it self-verify rendered output in a coding loop — it wrote, rendered, spotted layout bugs, and fixed them autonomously in ZCode. Professional-work scores on GDPval-AA v2 and OfficeQA Pro lead all compared models, including Opus 4.8 and GPT-5.6 Terra, though pure visual-perception benchmarks remain a relative weakness.

Takeaways
GLM-5.3-Flash scores 57 on the Artificial Analysis intelligence index, matching Claude Opus 4.8.
Promotional pricing is 4.5 cents per task — roughly 1/40th the cost of Opus 4.8.
The 320B-parameter model activates only 18B parameters across 45 layers, using a hybrid sparse-attention and linear-attention architecture.
IndexPool compresses 4 indexer cache vectors into 1, cutting latency and memory overhead under a 1M-token context window.
Attention computation drops 3.01x and KV cache drops 4.44x compared to GLM-5.3.
On DeepSWE v1.1 it scores 63.4 vs. Opus 4.8's 58.0; on AutomationBench, 48.8 vs. 41.0.
All pre-launch anonymous traffic on OpenCode and OpenRouter was served by a domestic Chinese chip cluster with custom inference-engine optimizations yielding a 3x performance gain.
The model is released under the MIT license with weights on HuggingFace, compatible with SGLang, vLLM, and TokenSpeed.
Native multimodality enables self-verifying coding loops: the model renders its own output, detects visual bugs, and fixes them.
Professional-work benchmarks GDPval-AA v2 (1773) and OfficeQA Pro (62.4) lead all compared models, including Opus 4.8 and GPT-5.6 Terra.
Pure visual-perception scores on benchmarks like BabyVision still trail GPT-5.6 Terra and Gemini 3.7 Flash.
Conclusions

A 40x price gap between models with equivalent benchmark intelligence suggests the premium on frontier API access has been a function of market position, not inference cost.

Running a full week of anonymous frontier-model traffic entirely on domestic Chinese chips — and achieving NVIDIA-comparable per-token costs — is a concrete counterpoint to the assumption that cutting-edge inference requires Western silicon.

Releasing a 320B model under MIT rather than a restricted or custom license removes the usual compliance friction that keeps Western enterprises from adopting Chinese open-weight models.

The hybrid attention architecture (sparse + linear) is not a research curiosity; it is the direct mechanism that makes 1/40th pricing possible at scale, and it is now publicly reproducible.

Native multimodality closing the loop from code generation to visual self-verification shifts the agent workflow from 'generate, then human checks' to 'generate, render, self-correct' — a meaningful reduction in the human-in-the-loop tax.

Concepts & terms
Hybrid Attention (Sparse + Linear)
An attention architecture combining linear attention (which captures local dependencies via recurrence) with sparse attention (which uses a lightweight indexer to recall global context). This hybrid approach structurally reduces the computational cost of long-context inference compared to full dense attention.
IndexPool
A weighted pooling technique that compresses an indexer's 4 cache vectors into 1, specifically designed to reduce latency and memory overhead when handling 1M-token context windows in the sparse-attention pathway.
mHC (Manifold Constrained Hyperconnection)
A technique that constrains hyperconnections in the model to lie on a low-dimensional manifold, improving scaling efficiency — the model reaches higher performance for a given parameter and compute budget.
EPD (Encode–Prefill–Decode) Disaggregation
A production inference architecture that separates the encode, prefill, and decode phases of transformer inference onto different hardware resources, allowing each phase to scale independently and improving throughput under load.
W8A8 Quantization
A quantization scheme where both weights (W) and activations (A) are stored in 8-bit integer format, reducing memory footprint and accelerating inference with minimal accuracy loss compared to higher-precision formats.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗