Model · Z.ai · Text
GLM-5.3-Flash
Z.ai's efficient GLM-5.3 tier under plain MIT — a 320B sparse model with 18B active, natively multimodal on input, at roughly a tenth of frontier pricing. The most permissively licensed model at this capability level.
- Modality
- Text
- License
- MIT (Open)
- Parameter size
- 320B total, 18B active
- Context window
- 1,000,000 tokens 1M tokens with 128k max output. Hybrid sparse + linear attention with Manifold-Constrained Hyper-Connections — the first hybrid architecture in the GLM series. Note the published evaluations were run at 300k context, not the full 1M the API advertises.
- Released
- August 26, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- vision, reasoning, code
What you need to run it
- Mac
- Yes. Any Apple Silicon Mac (M1 or newer) with at least 256 GB of memory. Newer chips are faster, but memory is what decides whether it runs.
- Windows / Linux PC
- Server only. Needs a server with several graphics cards, about 320 GB (640 GB recommended) of video memory in total — not a desktop PC.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
Fits only a 256 GB Mac Studio (Ultra chip) or a multi-GPU server; a 192 GB Mac is just too small.
Numbers are for: 4-bit MLX build of the 320B model (about 204 GB); the maker's FP8 weights are about 328 GB. Our estimate from the model’s size — the maker publishes no figure.
Strengths
- Plain MIT licence — genuinely unencumbered, unusual at this capability level
- Outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price; Terminal-Bench 2.1 84.3
- Natively multimodal input — video, image, text and files — so it can observe interfaces and rendering results, closing the loop across code, browsers and GUIs
- Hybrid attention sharply reduces long-context serving cost while preserving long-context accuracy
- $0.15 input / $0.50 output per MTok at list price
Weaknesses
- Reasoning cannot be disabled — thinking.type only accepts 'enabled'
- Multimodal in, text only out — the vision encoder does not mean image generation
- The reasoning_effort ladder is low / high / max, with no medium step
- Expert count and full architecture detail are undisclosed
- 1M context is an API capability, not an evaluated one — the model card's benchmarks stop at 300k
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | MIT licence |
| Z.ai API | hosted-api | API key required |
Official sources
- Model docs docs
- Model card model
- Pricing docs
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: size of the Hugging Face weights and mlx-community/GLM-5.3-Flash-4bit.
- — Verified current — still under Latest Models on Z.ai's pricing page at $0.15 in / $0.03 cached / $0.50 out per MTok, matching the list price recorded here. No newer GLM Flash model is listed.
- — Initial entry — first Z.ai model in the catalog. Released 2026-08-26, eight days after the flagship GLM-5.3 (2026-08-18). The Flash variant is catalogued rather than the flagship because it is plain MIT, while GLM-5.3 carries a custom licence requiring a Z.ai security review for MaaS operators above $10B revenue. Pricing recorded at list ($0.15/$0.50); a 50% promotional rate expires 2026-09-09. Release date taken from Z.ai's release notes, NOT from the arXiv paper linked on the model card — that paper (2026-02-17) is the shared GLM-5 base-model report, not this model's release.
Esc