Zhipu open-sources GLM-5.3-Flash: matches Claude Opus 4.8 at 1/40 the price
智谱 · wechat · 2026-08-26
Zhipu has released and open-sourced GLM-5.3-Flash (320B-A18B), the first natively multimodal model in the GLM-5 family. It scores 57 on the Artificial Analysis Intelligence Index, entering the global frontier band and matching Claude Opus 4.8, with comparable coding performance on Z.ai CodeBench — at 1/40 of Opus pricing (1/20 of GLM-5.3 during a limited discount). It beats the twice-as-large GLM-5.2 across benchmarks.
Key points:
- Architecture: first open-source frontier model combining sparse and linear attention, plus Manifold-Constrained Hyper-Connections (mHC) for better scaling; trained on 30T tokens of multimodal data. Attention compute and KV cache drop 3.01x and 4.44x vs GLM-5.3.
- Visual coding: natively integrated vision lets the model inspect and iterate on its own outputs; it autonomously built a 400㎡ chef's kitchen in Blender over 16 hours, and Browser/Computer Use Agents verify rendering and interaction.
- Domestic chips: all traffic was served on Chinese accelerator clusters; a custom SGLang-based engine with W8A8 quantization, mixed cache quantization and an Encode–Prefill–Decode disaggregated architecture delivered 3x end-to-end speedup, reaching per-token costs on par with mainstream NVIDIA GPUs.
- Availability: API, GLM Coding Plan (10,000 daily trial cards) and open weights are live. Pre-release, it was tested anonymously as Ox-Alpha on OpenCode and OpenRouter, becoming the most popular model of the week.
Related event: Zhipu Open-Sources GLM-5.3-Flash: 320B MoE Multimodal at 1/10th the Cost(48 posts)→
More from Infra
- Copper to optical: optical backplanes seen as necessary as racks move to optics — BenBajarin · 2026-10-16
- Polymarket prices 28% odds a US state enacts data center moratorium by year-end — Polymarket · 2026-10-16
- Pennsylvania residents reject $10,000-per-household offer for massive Amazon data center — Polymarket · 2026-10-16
- PyTorch now runs natively on AWS Trainium with zero code changes — PyTorch · 2026-10-16
- Chinese labs run on ~1/10 US compute; ByteDance pays 2-4x to hoover up SEA supply — robleclerc · 2026-10-16
- Zymtrace raises $12.2M and hires ML engineers to build agents that fix GPU bottlenecks — cfregly · 2026-10-16