Z.ai's GLM-5.3-Flash: 320B open-weight MoE with 18B active and a 1M-token context

What Z.ai released

Z.ai released GLM-5.3-Flash on August 26, 2026: an open-weight Mixture-of-Experts model with 320 billion total parameters, of which only around 18 billion activate for any given token. According to a dev.to breakdown, it is the first natively multimodal model in the GLM-5 series — it takes text, images, video and files as input — and it holds a one-million-token context with up to 128K tokens of output. The trained weights sit on Hugging Face under an MIT license.

The Flash variant arrived twelve days after the full-size GLM-5.3, and it is deliberately not the flagship. Dev.to characterises it as a play for the best capability at the lowest cost rather than a bid to top any single leaderboard.

How a 320B model runs like an 18B one

The MoE layout divides the model into specialist sub-networks, and a router wakes up only a fraction of them per token. The network keeps the capacity of a 320B model while each token costs roughly what an 18B model would cost to process.

Z.ai's documentation, as summarised by dev.to, adds a hybrid attention system: sparse attention for the most relevant long-range context paired with linear attention for local dependencies, which the company describes as a first for an open frontier model. Reported effects include 3.0x less attention computation and a 4.4x smaller KV cache than the full GLM-5.3. A component named IndexPool compresses indexer key vectors, which dev.to credits as a big reason a million-token context can be served economically at all. The model also uses what Z.ai calls Manifold-Constrained Hyper-Connections to improve scaling efficiency, and it was pretrained on a 30-trillion-token multimodal corpus.

The numbers, with a caveat

Z.ai's launch figures, relayed by dev.to, put GLM-5.3-Flash at 48.8% on AutomationBench (up from 26.2% for GLM-5.2), 63.4% on DeepSWE 1.1 (up from 46.2%), 56.3% on NL2Repo-Bench and 78.4% on Toolathlon-Verified. The first two measure agentic workflows and end-to-end software engineering; the last two measure natural-language-to-repo generation and tool use. Dev.to is explicit that these are vendor-published launch numbers rather than independent results.

On the Artificial Analysis Intelligence Index the model scores 57 at a discounted cost of about $0.045 per task, a capability level Z.ai says cost roughly ten times more months earlier. Third-party hosts charge around $0.15 per million input tokens and $0.50 per million output tokens, figures dev.to verified in early September 2026 and flags as likely to change.

A stealth launch on domestic silicon

Before the reveal, Z.ai reportedly ran the model anonymously under the codename ox-alpha on OpenRouter and OpenCode, where it became the most-used model of that week before anyone knew who had built it. The more consequential detail is that all of that traffic was served on Chinese-made AI chips rather than Nvidia hardware, using a custom inference engine built for the hybrid architecture, at speeds dev.to reports came close to mainstream Nvidia setups. For builders, that matters less as geopolitics than as supply: more viable serving hardware means more inference capacity, which is one of the forces pushing per-token prices down.

Vision inside the loop

Because vision is built in rather than bolted on, the model can inspect its own output — rendered pages, screenshots, charts, documents and video — inside the coding loop, and catch visual problems a text-only model would miss. Dev.to lists visual debugging from a broken-UI screenshot, document-to-spreadsheet pipelines across the 1M window, long-video analysis and design-to-code as the practical payoffs.

What builders should note

API users can treat it as a cheap default tier and benchmark it against their current workhorse for cents per failed experiment. Self-hosters get the MIT license and public weights, but the 18B-active figure is misleading on hardware requirements: the full 320B of weights still has to fit in memory. It is also open-weight rather than open source in the strict sense — training data and training code are not released.

Why it matters

GLM-5.3-Flash bundles three shifts into one release. It shows near-frontier coding and agentic behaviour at commodity prices, scoring 57 on the Artificial Analysis index at a few cents per task. Its architecture — MoE sparsity plus hybrid attention plus KV compression — makes a million-token context servable rather than theoretical. And its stealth launch demonstrated that a top open model can run at production scale on non-Nvidia silicon. The AutomationBench jump from 26.2 to 48.8 in a single generation is the figure to watch, because unsupervised multi-step agents are what actually consume budgets. If independent testing confirms Z.ai's numbers, the floor for what an open, cheap, self-hostable model must deliver has just risen.