GLM 5.3: What's New, Benchmarks, and How to Access It (2026)

Z.ai shipped GLM 5.3 without touching the base model. Here's what actually changed, what's still vendor-reported, and how to run it today.

GLM 5.3 released on August 14, 2026, and the headline is what Z.ai didn't do: no new architecture, no retraining, same base model as GLM 5.2. Every gain comes from scaled post-training, what Z.ai calls environment scaling, running the model through more and more diverse task environments. Their own summary: "Scaling post-training is all we did for GLM-5.3."

The claimed results from that one lever are large enough to raise eyebrows, which is exactly why the vendor-reported caveat matters more than usual here. Here's the current picture.

What changed from GLM 5.2

The numbers below are Z.ai's own, from the launch materials. Nothing has been independently replicated yet.

The pattern in the table is consistent: the biggest jumps are on long-horizon work, multi-step tasks that run for hours and punish a model that loses the thread. That matches the positioning, GLM 5.3 as a model for complex software engineering and agent workflows rather than a general chat upgrade. If your workload is short prompts and single answers, 5.2 and 5.3 will feel similar. If it's agents running long tool chains, the delta is the whole point.

A six-fold jump on Terminal-Bench from post-training alone is a remarkable claim, and remarkable claims earn scrutiny. Watch the independent leaderboards over the next couple of weeks, and validate on your own workload before treating the table as settled.

What carried over from 5.2

Because the base model is unchanged, the spec sheet mostly reads like GLM 5.2's. Z.ai positions the 5.2 base around a usable 1M-token context, and 5.3 inherits that; a separate 5.3 model card with its own numbers hadn't been published at the time of writing, so treat context and parameter details as carried over rather than re-confirmed. The serving story carries over the same way. The 5.2 line already runs on vLLM and SGLang, and our SGLang deployment guide for GLM 5.2 describes the stack that 5.3's weights should slot into when they arrive.

The same-base decision also explains the release cadence. Z.ai shipped 5.1 in the spring, 5.2 in early summer, and now 5.3, each one leaning harder on post-training while the architecture holds still. If the 5.3 benchmark table survives independent testing, it says the frontier labs' post-training gap is closing faster than the architecture gap, which is the argument Chinese open-weight labs have been winning all year.

The security capability nobody planned for

The strangest part of the launch: Z.ai reports the model developed offensive security skills faster than expected, moving from isolated bug-finding to what they describe as "reasoning across multiple stages of exploitation, forming coherent plans for complete exploitation chains."

The concrete side of that claim is a public disclosure ledger: 2,436 vulnerabilities reported across 269 open-source projects since GLM 5.2, 1,097 of them rated critical or high severity, with 53 CVEs assigned at launch. Whatever you make of the benchmark table, the CVE ledger is externally checkable, which makes it the most credible evidence in the release.

For infrastructure teams the practical read is simpler: the same long-horizon reasoning that finds exploit chains is what you want for multi-hour agent runs, and security research happens to be a measurable proxy for it.

The weights arrived, attached to a surprise: GLM 5.3 Flash

On August 26 the open-weights story resolved in an unexpected shape. Instead of the flagship, Z.ai released GLM 5.3 Flash: a 320B-parameter MoE with 18B active per token, natively multimodal (the first in the GLM-5 series), a 1M-token context window, and full weights on Hugging Face under MIT. It had spent the previous week running anonymously as "Ox Alpha" on OpenCode and OpenRouter, served on Chinese-made chips. The reveal came with a twist worth noticing: Z.ai's own tables show Flash beating the flagship 5.3 on several agent benchmarks at roughly a tenth of the price, the same Flash-embarrasses-the-flagship pattern DeepSeek went through this month; the full GLM 5.3 vs GLM 5.3 Flash comparison untangles how much of that holds up. The flagship 5.3's weights remain unreleased, so self-hosting the model this post covers is still not possible. Flash, though, ships with day-one vLLM and SGLang support and is very much self-hostable, in the multi-GPU sense; the GLM 5.3 Flash hardware requirements breakdown has the full math.

How to access GLM 5.3 today

Two real routes right now.

Z.ai's own API and GLM Coding Plan serve 5.3 directly; check their current pricing before budgeting, since plan tiers moved around at launch.

Or run it through Yotta AI Gateway, where GLM 5.3 is already live behind one OpenAI-compatible API, alongside GLM 5.2, 5.1, DeepSeek V4, Qwen 3.8, Kimi K3, and the rest of the catalog. If you're already on 5.2 through the Gateway, trying 5.3 is a model-string change, which is the lowest-effort way to test whether the long-horizon gains show up on your workload.

Frequently asked questions

What is GLM 5.3's context window?

Z.ai hadn't published a standalone 5.3 spec sheet at the time of writing. It shares GLM 5.2's base, which Z.ai positions at a usable 1M-token context, so expect the same until the model card says otherwise.

What does GLM 5.3 cost?

Z.ai serves it through its API and the GLM Coding Plan; rates moved around at launch, so check their current pricing page before budgeting. It's also live on Yotta AI Gateway, where pricing is flat and posted in the console.

When was GLM 5.3 released?

August 14, 2026, on Z.ai's API and GLM Coding Plan. GLM 5.3 Flash followed on August 26 with open weights; the flagship's weights are still pending.

Is GLM 5.3 open source?

The flagship, not yet. But GLM 5.3 Flash, released August 26, is fully open: MIT weights on Hugging Face, 320B total parameters with 18B active. The flagship 5.3 remains API-only for now.

What's the difference between GLM 5.3 and GLM 5.2?

Same base model. All of 5.3's gains come from scaled post-training, and the claimed improvements concentrate on long-horizon tasks: Terminal-Bench 3.0 jumps from 4.6 to 28.3 and DeepSWE from 46.2 to 66.9, per Z.ai's own numbers.

Is GLM 5.3 better than DeepSeek V4?

No independent head-to-head exists yet, and both launch benchmark tables are vendor-reported. They target different sweet spots: DeepSeek V4 leads on price aggression, GLM 5.3 on the long-horizon agent story. Both are live on Yotta AI Gateway, so testing them head-to-head on your own workload is two model strings.

Can I self-host GLM 5.3?

The flagship, no, its weights are unreleased. GLM 5.3 Flash, yes: MIT weights with day-one vLLM and SGLang support, though a 320B MoE is an 8-GPU-node commitment. The GLM 5.3 Flash hardware requirements breakdown has the full math, and our GLM 5.2 SGLang guide covers the serving stack the GLM line runs on.

What is GLM 5.3 Flash?

A smaller open-weight sibling released August 26: 320B total parameters, 18B active, natively multimodal with image and video input, 1M-token context, MIT license. It previously ran in stealth as "Ox Alpha." Z.ai's launch pricing is $0.15 per million input tokens and $0.50 output, and its own benchmarks show it matching or beating the flagship on several agent tasks. The flagship vs Flash comparison covers which to use.

Is GLM 5.3 on Yotta?

Yes. GLM 5.3 is live on Yotta AI Gateway with an OpenAI-compatible API, alongside the rest of the GLM line.

Bottom line

GLM 5.3 is a bet that post-training scale beats architecture churn, and the claimed returns are big enough that someone will verify them soon. The CVE ledger gives the long-horizon story more external evidence than most launch tables get. The open question is how much of the benchmark jump survives independent testing.

The practical move while that settles: if you run agents or long coding tasks, put 5.3 next to 5.2 on your own workload and measure. GLM 5.3 and 5.2 are both live on Yotta AI Gateway, one API key and a model string apart. The weights dropped in Flash form on August 26; the flagship's are still pending, and this post tracks both.