GLM-5.3-Flash: the 320B That Fits Because Its Attention Does Not Grow

Local AI Updated 2026-11-04 · 3 min read · #glm #moe #linear-attention #sparse-attention #local

Status board (checked 2026-11-04)


Who made it	Zhipu / Z. AI (Beijing)
License	MIT (model card on HuggingFace: `zai-org/GLM-5.3-Flash`)
Export-control status	⚠️ China catalogue draft (comment period) — registered for overseas export; watch the date box

The numbers


Total / active	~320B total, ~18B active per token — 45 layers: 34 linear-attention + 11 sparse-attention
Why it fits at all	the hybrid attention cut attention compute ~3× and KV cache to a third vs GLM-5.3 — context is affordable now (up to 1M marketed)
Maya pack sizes	XS 73.4 GB (64 GB RAM + 24 GB card at 128K) · S24 94.7 (cards ≤24 GB, +14% decode) · S 96.5 (default) · M 116 · L 156.3
Measured speeds	148.6 tok/s on 2×5090+4090 at 32K · 122 on 2×H100 at 1M · 118 on 4×4090 · 81 on one RTX PRO 6000 · 16.6 on a lone RTX 5070 12 GB + 4-channel DDR4
Quality kept	97.7–99.2% of the FP8 model’s zero-shot accuracy; Maya-L picks the same next token 90% of the time
Trap	“18B active” ≠ runs in 20 GB of fast memory: the whole expert table must stay addressable (RAM + SSD)

Getting started (3 steps)

Clone Project Maya and run its setup — it recommends a tier for your machine (or pass --model glm-maya-s24 to pick one); Strata users get it through strata launch’s model list.
A 64 GB RAM + 24 GB card box runs Maya-XS at 128K context; 32 GB rigs can start the smallest tiers and let the SSD carry the rest.
Keep context modest at first, then raise it — the whole point of the architecture is that context costs a third of what it used to. Up to 8 requests can be answered at once for agent work.

Best use

Agentic coding and long-document work at home: the frontier-adjacent pick when gpt-oss-class models aren’t enough but datacenter money isn’t available. The M tier is calibrated toward tool calls and front-end code; a community “Derisked” edit trims blanket refusals.

Honest weaknesses

Heavier than the trend’s face: the comfortable rigs start at 64 GB+ RAM and a big NVMe, and a slow SSD hurts every number.
New architecture = young tooling: some engines still caught up on Day-1 features; the popular DFlash2 drafter is CC BY-NC-ND (non-commercial).

FAQ

Why do people say it is “cheaper than DeepSeek at long context”?

The linear+sparse hybrid shrinks KV growth — 1M tokens is the ceiling, but even at 128K home context the memory bill is far smaller than older 300B-class models.

Is it multimodal?

Yes, natively (text+images); the picture encoder ships with the packs and routes to CPU or GPU by card size.

Does the China export story kill it overseas?

Not today: MIT-licensed weights sit on HuggingFace worldwide and Chinese labs completed overseas export registrations. The draft rules add paperwork noise, not a wall — but watch the date on this box.

📌 Summarised from official docs and hands-on samples. Product features and pricing change often — official site wins.

Older: How a 100B+ Model Fits in a Gaming PC — what Strata and Project Maya changed →