GLM-5.3-Flash

// Overview

GLM-5.3-Flash is the Flash-tier model Z.ai released on August 26, 2026, and the first model in the GLM-5 series with vision built into the model rather than bolted on. It takes text, images, video and files in and returns text, carries a 1M-token context window, and runs in thinking mode only — Z.ai's documentation states that `thinking.type` supports `enabled` and that thinking cannot be disabled. A `reasoning_effort` parameter accepts `low`, `high` and `max`, defaulting to `max`.

The headline is intelligence per dollar rather than a new ceiling. Z.ai reports GLM-5.3-Flash scoring 57 on the Artificial Analysis Intelligence Index v4.1.1 at $0.045 per task at the discounted rate, and says it outperforms GLM-5.2 across six coding and agentic benchmarks at roughly one tenth of GLM-5.2's price — 63.4 against 46.2 on DeepSWE v1.1 and 48.8 against 26.2 on AutomationBench. On Z.ai's in-house Z.ai Code Bench v1.0, run in Claude Code 2.1.207, the model at max effort reaches 29.0 against the 29.5 Z.ai lists for Claude Opus 4.8.

The efficiency comes from the architecture. GLM-5.3-Flash pairs linear attention, which captures local dependencies through state modelling, with sparse attention that retrieves global context through a lightweight indexer. Z.ai reports that against GLM-5.3 this cuts attention compute by 3.01× and KV cache size by 4.44×, and notes that compared with the GLM-4.5 series the model keeps a similar total parameter count (320B against 355B) while nearly halving both the activated parameters (18B against 32B) and the layer count (45 against 92).

Native vision is aimed squarely at the coding loop. Z.ai's case is that for frontend work, game development and 3D simulation the deliverable is an interface or an experience rather than code, and many failures only surface on rendering or playtesting — so the model is trained to decide when to look, inspect its own rendered output and iterate. Z.ai's documented use cases run from screenshot-to-Next.js reproduction and Godot game prototypes to Blender scenes, build123d CAD models, PPTX/DOCX/XLSX office deliverables and computer-use workflows.

Before launch the model ran anonymously as "ox-alpha" on OpenCode and OpenRouter, where Z.ai says it became the most popular model of that week. Z.ai also states that all of this traffic was served on a large-scale cluster of Chinese AI chips, using a dedicated SGLang-based inference engine with W8A8 quantization, hybrid INT8/FP8/BF16 cache quantization and a disaggregated Encode–Prefill–Decode architecture, reaching a 3× end-to-end serving improvement over its own baseline on the same hardware.

The weights are public under the MIT licence in the `zai-org/GLM-5.3-Flash` Hugging Face repository, with a BF16 variant alongside it, and local deployment is documented for SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth. The hosted model is served through the Z.ai API as `glm-5.3-flash` and is included in the GLM Coding Plan with three times the quota of GLM-5.3.

// Benchmarks

This model's scores

1. DeepSWE v1.163.4%
2. AutomationBench48.8%
3. Z.ai Code Bench v1.0 (max effort)29%
4. Artificial Analysis Intelligence Index v4.1.157

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Z.ai's list price. A 50% launch discount halves it to $0.075 input, $0.015 cached input and $0.25 output until 24:00 on September 9, 2026 (UTC+8). Cached-input storage is listed as free for a limited time. The GLM Coding Plan bills separately on a points quota and grants three times the GLM-5.3 quota for this model.

Pricing source ↗

// Strengths

• DeepSWE v1.1 of 63.4 against 46.2 for GLM-5.2, at roughly a tenth of GLM-5.2's price
• AutomationBench of 48.8, nearly double GLM-5.2's 26.2
• Z.ai Code Bench v1.0 at max effort of 29.0, against the 29.5 Z.ai lists for Claude Opus 4.8
• Native multimodal input — images, video and files — used inside the coding loop rather than as a separate vision model
• 1M-token context at low serving cost: 3.01× less attention compute and a 4.44× smaller KV cache than GLM-5.3
• MIT-licensed weights with documented SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth deployment paths

// Best for

• Vision-driven frontend work: turning screenshots, page images or screen recordings into a running project, then comparing renders against the reference
• High-volume agentic coding where cost per task matters more than the last point of benchmark headroom
• Office and document deliverables — PPTX, PDF, DOCX and XLSX — with the model rendering and visually inspecting its own output
• Game, Blender 3D and CAD workflows that need repeated build → render → inspect → refine loops
• Computer-use and GUI automation in environments with no structured API
• Self-hosted deployment where MIT-licensed weights and a small activated-parameter count are the constraint

// How to access

// GLM Flash — every version

The full lineage of the GLM Flash line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.