GLM-5.3-Flash from Z.ai: MIT licence, 328 GB of weights, you build the rest

On 25 August 2026, at 06:43 UTC, the zai-org organisation — Z.ai, formerly Zhipu AI — published GLM-5.3-Flash on Hugging Face: the first natively multimodal model in the GLM-5 series, MIT licence, weights downloadable without authentication. For anyone assessing a model of this size to run inside their own perimeter, the question that matters is not whether the licence allows commercial use — here it does, with no thresholds or extra clauses — but how much it really costs to keep it running, and how much of what the vendor shows on its own platform stays outside the file you download. We verify it here, as of 27 August.

The facts, in order

Publication: repository created on 25 August 2026 at 06:43:14 UTC, last updated on 26 August at 16:31. Organisation zai-org, verification badge absent (isVerified: false) despite 152 published models and 19,252 followers — the check has to be made card by card, not by the account’s reputation.
Access: gated: false. Verified without authenticating: requesting a weight shard returns a direct redirect to the CDN (HTTP 302), not a refusal (401) — no agreement to accept before downloading.
Adoption, today: 766 likes, 0 downloads according to the public counter, read on 27 August — the download figure stays at zero even two days after publication, more likely a counter quirk than a sign of low interest, given the likes.
Weight: 321.3 billion total parameters per the repository’s metadata (the vendor rounds to 320 billion, with 18 billion active per token); 97.8% of parameters (314.4 billion) are natively FP8, the rest BF16 for precision-sensitive modules — normalisations, routing mechanisms, the indexer. On disk: 62 safetensors files, 328.3 GB — 305.8 GiB — in total.
Architecture: 45 layers, 34 of them linear-attention and 11 “DeepSeek Sparse Attention” — one in every four; 288 routed experts plus one shared, eight activated per token; a native vision encoder, built in rather than bolted on later. Context window declared in the configuration file: 1,048,576 tokens — the same figure already measured on the predecessor GLM-5.2, not an extension.
Base: none. No base_model tag in the repository; the card states the model “starts from a newly trained base model” — not a fine-tune, no upstream licence chain to check.
Licence: MIT, text verified (HTTP 200), with no additional attachments — no separate use-policy file: four common filenames for this kind of document (USE_POLICY.md, NOTICE, MODEL_LICENSE, ACCEPTABLE_USE_POLICY.md) all return 404. Here the declared licence is the one that applies, in full.

What it is, in practice

GLM-5.3-Flash is built, according to the title of the GLM-5 family’s technical report, for agentic use more than conversation: “GLM-5: from Vibe Coding to Agentic Engineering”. The hybrid architecture that defines it — linear attention alternating with sparse-attention layers, plus a connection technique the card calls Manifold-Constrained Hyper-Connections — serves one declared purpose: keeping the model’s capability while cutting the cost of serving it at long context. The vendor claims performance ahead of its own GLM-5.2 and close to Claude Opus 4.8 on coding and agentic tasks — figures read off a table, not verified by us: we have not run the model.

The sum the disk weight does not do

328.3 GB is the number you see first, and it is only the opening term of the sum. The vendor itself, in its own technical documentation, locates the real constraint not in compute but in memory: compared with its bigger sibling GLM-5.3, this version’s hybrid architecture “reduces attention computation and KV cache size by 3.01× and 4.44×, respectively” — a saving that only makes sense once you start from a context close to a million tokens, where the attention cache can weigh as much as the weights themselves. To serve it at its own scale, Z.ai states it built “a dedicated inference engine for this architecture on top of SGLang”, with tensor parallelism for linear attention, hybrid INT8/FP8/BF16 cache quantisation and a disaggregated architecture that splits encoding, prefill and decoding across independently scheduled pools of machines — deployed, again by the vendor’s own account, across “tens of thousands of domestically developed accelerators”. Downloadable weights without the compute that makes them runnable are not sovereignty: here the missing sum is not just the memory to load 328 GB, it is the systems engineering the vendor had to build to serve the same model at its own scale — and that the card does not estimate anywhere for whoever installs it in-house.

Weights downloaded, platform not included

Whoever downloads these weights gets a language-and-vision model: it generates text, reads images and video, nothing else pre-built. Z.ai’s hosted-platform documentation lists, alongside the bare model, an integrated web search, dedicated agents for slides, translation and video editing, and a persistent context cache for long conversations: application infrastructure built on top of the checkpoint, not contained in the safetensors files you download. The million-token window is available in both cases — it sits in the configuration file, not as a service exclusive — but orchestrating tools, conversation memory and agents around the model remains work you redo yourself, with a free licence and no guidance on how much compute it takes to do so at that scale. We have already written about the gap between what a card declares and what it leaves out: here it is not a security chapter that is missing, it is the infrastructure estimate — a gap just as concrete for whoever has to put the model into production.

What we do not know

We have not run the model: no claim of ours about quality, safety or real-world behaviour. The comparisons with GLM-5.2 and Claude Opus 4.8 are the vendor’s own table figures, not an independent trial. The technical report’s abstract (arXiv 2602.15763) specifies no provenance, no licences and no composition for the training corpus declared at 30 trillion multimodal tokens: we know the aggregate size, not the source. We offer no legal qualification of the MIT licence. And we do not know the minimum hardware for a deployment smaller than the huge one Z.ai describes for its own scale: no GPU card, no memory threshold is given for anyone who does not need to serve millions of requests a day, as already observed on another open-weight model published the week before.

How we solve it

For a model this size, verification does not end at the download: it enters the model registry as a dated line — version and file digest, acquisition channel and date, the MIT licence text archived exactly as it read on 27 August, the approved uses and who in the organisation approved them — ready to produce at an inspection, for a client during a tender, or for a board. The same system holds together models, data, contracts, archives and documents in a single operating model, on which AI agents execute decisions with an operator in command: for large enterprises, defence, public administration and healthcare. Always on-premises on autonomous machines that require no deep integration into the client’s network, or dedicated cloud with a dedicated VPN and a data centre in Italy — always with shared management: you do not need to already have, in-house, someone who administers these systems.

See the service · Talk to an engineer

Have you already downloaded, or are about to download, 328 GB of open weights, with no one in the organisation able to say how much cache, how much memory and what inference engine it takes to bring them into production? Half an hour with one of our experts, at no cost: we bring the real estimate of infrastructure, licence and declared data before it ends up in a tender — the boxes on who approved it and on what hardware stay, for now, to be filled in.