All models

Page last updated on

Zhipu AI logo

GLM-5.3

By Zhipu AI

MoE1M contextAgentic codingCyber Unannounced license

The Z.ai coding flagship, rebuilt through post-training alone on the same 753B base as GLM-5.2. Open-source SOTA on Terminal Bench 3.0 and state of the art on CyberGym. The weights are not out yet.

Release Blog

Specifications

Total parameters
753B (40B active)
Active parameters
40B
Architecture
MoE
Architecture class
GlmMoeDsaForCausalLM
Context window
1M
Vocab size
154,880
Modality
Text
License
Unannounced
Released
August 2026
Best for
Agentic coding, long-horizon engineering tasks, and vulnerability research

Good to know

  • The weights are not released yet. Z.ai shipped GLM-5.3 on August 14, 2026 through the GLM Coding Plan and ZCode, and wrote that the weights follow 'in two weeks after launch, once safety evaluation and hardening are complete.' That points to roughly August 28, 2026.

  • Thinking cannot be disabled. GLM-5.3 takes a `reasoning_effort` of low, high, or max, and max is the default and the recommended setting for coding. If your application currently sends `thinking.type: "disabled"`, switch it to enabled and set `reasoning_effort` to low before moving to the glm-5.3 model ID, or the request fails.

  • Z.ai reports that GLM-5.3 found 2,436 vulnerabilities across 269 open source projects when run against real codebases with security teams in China, including 1,097 rated medium to high. As of the announcement, 53 were publicly disclosed and 2,383 remained under embargo. The oldest flaw dated to 1981, and the average vulnerability had survived 26.6 years before discovery. Z.ai tracks the disclosure process in a public ledger.

Architecture

GLM-5.3 is a deliberate controlled experiment. Z.ai kept the base model, the architecture, and both parameter counts identical to GLM-5.2, then spent a month scaling post-training alone.

The stack underneath is the one GLM-5.2 introduced. IndexShare handles long-context efficiency by reusing one sparse-attention indexer across each group of four layers, which keeps the KV cache affordable at 1M tokens. SAO carries the reinforcement learning on long-horizon tasks. Training runs on slime, the open source RL framework from Z.ai, with Megatron on the training side and SGLang on rollout.

What changed is scale and plumbing:

  • More environments, not more parameters. Z.ai built pipelines that synthesize long-horizon task environments end to end, along with the reward signal for a subset. A judge agent attempts each task to confirm it is solvable, and verifiers are written without access to the reference solution.
  • A faster training loop. Workload-aware scheduling and a local-storage caching layer lifted end-to-end RL throughput by more than 2.3x on long-horizon coding tasks. Training and rollout log probabilities were aligned to the 1e-7 level.
  • Fewer tokens per task. The efficiency gain is the underrated part. At max effort GLM-5.3 reaches 34.5% on the in-house Z.ai Code Bench using about 75K output tokens per task, where GLM-5.2 needed 96K to reach 23.4%.

For more models like this one, browse the full open source LLM directory, compare the field in the best open source LLMs in 2026, or read the open source LLM ecosystem statistics.

Benchmarks

GLM-5.3 is the same base model as GLM-5.2 with a month of extra post-training on top, so the honest comparison is the GLM-5.2 column.

GLM-5.3 vs frontier models

Scores published by Z.ai in the GLM-5.3 announcement. Higher is better on every row. Scores marked n/a were not reported. ExploitGym counts tasks completed under time-normalized budgets, so it is split into the 2 hour and 6 hour rows and plotted on a task-count axis. GDPval-AA v2 is an Elo rating from Artificial Analysis. Z.ai ran most agentic rows through the Claude Code 2.1.207 harness at max reasoning effort, so these measure the model plus that scaffold.

GLM-5.2 Kimi K3 DS-V4-Pro-0813 Qwen3.8-Max Opus 4.8 Fable 5 GPT-5.6 Sol GLM-5.3

Each model keeps the same color across every benchmark, stacked top to bottom in this order. Hover or focus a row to read all six scores.

Terminal Bench 2.1
88.2 #3 of 8
Terminal Bench 3.0
28.3 #3 of 6
DeepSWE v1.1
66.9 #4 of 8
NL2Repo
58.0 #3 of 6
ProgramBench Almost Solved
19.0 #3 of 7
FrontierSWE
78.1 #2 of 4
SWE-Marathon v1.1
42.5 #3 of 6
PostTrainBench
39.8 #2 of 6
View all scores as a table
Benchmark GLM-5.2Kimi K3DS-V4-Pro-0813Qwen3.8-MaxOpus 4.8Fable 5GPT-5.6 SolGLM-5.3
Coding
Terminal Bench 2.1 81.0 88.3 87.9 86.6 85.0 88.0 88.8 88.2
Terminal Bench 3.0 4.6 17.4 n/a n/a 21.1 33.7 34.6 28.3
DeepSWE v1.1 46.2 67.5 62.7 56.6 58.0 69.7 72.7 66.9
NL2Repo 48.9 58.0 61.1 55.9 69.7 n/a n/a 58.0
ProgramBench (Almost Solved) 9.5 17.5 n/a 10.5 15.5 33.0 23.0 19.0
FrontierSWE 67.5 n/a n/a n/a 66.5 88.2 n/a 78.1
SWE-Marathon v1.1 19.4 48.1 n/a n/a 48.8 33.1 42.5 42.5
PostTrainBench 31.7 32.0 n/a n/a 32.9 41.8 36.2 39.8
Cyber
CyberGym 77.2 80.0 83.3 78.5 78.1 83.8 83.6 84.5
ExploitBench 24.4 32.2 n/a 28.8 40.0 78.0 76.5 54.4
ExploitGym 2h (Tasks) 29 36 n/a 14 80 181 216 105
ExploitGym 6h (Tasks) 39 70 n/a 26 120 247 293 130
Agentic
Toolathlon Verified 59.9 76.5 74.1 72.5 76.2 74.7 74.9 73.0
AutomationBench v1.0.6 26.2 46.7 43.2 39.8 41.0 46.2 45.8 48.2
Agents' Last Exam (ALE-CLI) 23.8 27.6 25.7 27.0 25.7 23.8 28.6 28.5
HLE w/ Tools 54.7 59.8 60.0 56.2 57.9 63.9 64.5 62.5
GDPval-AA v2 (Elo) 1,508 1,682 1,590 1,739 1,588 1,743 1,730 1,769

Source: the GLM-5.3 model card. Best score in each row is marked.

Third-party evaluations

Independent scoring landed while the weights were still pending, so the numbers below all describe the hosted model. Both agree on the shape of the result: GLM-5.3 moved up a tier from GLM-5.2 and now sits level with the strongest open-weight competition.

Two Artificial Analysis charts. The top bar chart is the Intelligence Index, which combines nine evaluations. Claude Opus 5 (max) leads at 63, Claude Fable 5 (with fallback) 62, GPT-5.6 Sol (max) 61, Kimi K3 (max) 60. GLM-5.3 (max) is marked with an arrow at 60, tied with Kimi K3, ahead of Qwen3.8 Max at 58 and well ahead of GLM-5.2 (max) at 53. The bottom scatter plot charts Intelligence Index against cost per task on a log scale, with GLM-5.3 (max) arrowed onto the Pareto line at roughly $0.65 per task.
Artificial Analysis Intelligence Index. On the Artificial Analysis Intelligence Index, GLM-5.3 scores 60. That is a 7 point jump over GLM-5.2 at 53, from a model with an identical base and parameter count. It ties Kimi K3 (max) and lands one point behind GPT-5.6 Sol (max) at 61, with only Claude Fable 5 at 62 and Claude Opus 5 at 63 ahead. The scatter plot puts GLM-5.3 on the Pareto line at roughly $0.65 per task, to the left of Kimi K3 at a similar score. Artificial Analysis · Aug 2026
Arena Code Arena WebDev Pareto frontier scatter plot of Arena Score against blended price per 1M tokens on a log scale. A tooltip labels glm-5.3-max, Z.ai, MIT, at a score of 1,597 and $3.65 per 1M tokens, sitting on the green frontier line. Other labelled frontier points are claude-opus-5-max near $20, kimi-k3-max, qwen3.8-max, gemini-3.7-flash-high, deepseek-v4-flash-high, hy3, mimo-v2.5, solar-pro4, and granite-4.1-8b.
Code Arena WebDev Pareto frontier. On the Arena Code Arena WebDev board, glm-5.3-max scores 1,597 and holds a place on the Pareto frontier at $3.65 per million blended tokens. The frontier runs down from claude-opus-5-max near $20 through kimi-k3-max and qwen3.8-max to the GLM point. One detail worth flagging: Arena already labels the entry MIT, which reflects the expectation set by GLM-5.2 rather than a license Z.ai has announced for GLM-5.3. Arena · Aug 2026

API pricing

GLM-5.3 ships at the same API rates as GLM-5.2: $1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens. A capability jump at a flat price is the headline here. Z.ai lists cached input storage as free for a limited time.

Model Input · cache hit Input · cache miss Output
GLM-5.3 $0.26 $1.40 $4.40
GLM-5.2 $0.26 $1.40 $4.40

Per 1M tokens, as of August 21, 2026. Official pricing.

Against the closed frontier the gap is wide. Claude Opus 5 costs $25 per million output tokens, so GLM-5.3 is about 82% cheaper. GPT-5.6 Sol costs $30, a roughly 85% gap. The comparison that matters more is token efficiency, because GLM-5.3 spends fewer tokens to get further. On the in-house Z.ai Code Bench, the model reaches 34.5% at around 75K output tokens per task at max effort, against 23.4% at 96K for GLM-5.2. At high effort it reaches 31.4% at about 50K output tokens, above Claude Opus 4.8 at 29.5% with 120K. Claude Fable 5 still leads outright at 39.5%.

Most users will reach GLM-5.3 through the GLM Coding Plan rather than per-token billing. That plan moved to a points quota, counted separately for input, cached input, and output. Calls outside peak hours consume half the standard points. Peak is 14:00 to 18:00 UTC+8, Monday through Friday, so every other hour including the whole weekend gets the off-peak rate.

Sources: Z.ai pricing · GLM-5.3 announcement · Claude pricing · OpenAI pricing

What people are saying

“GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal... What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling.”

Jie Tang

@jietang · Founder and chief scientist, Zhipu AI

X Aug 19, 2026

“GLM 5.3 is finally here I had the chance to test it 24 hours before its official release, and what I noticed > it stays on track longer > handles complex coding tasks with fewer corrections > feels noticeably more reliable across multi step workflows..”

JUMPERZ

@jumperz

X Aug 14, 2026

“GLM 5.3 is actually a huge deal. They didn't make the underlying model bigger or fundamentally change the architecture. They just took GLM 5.2 and post trained it harder. Yet, that alone produced major gains, especially in complex coding, long-horizon agent tasks, and...”

AI Search

@aisearchio

X Aug 14, 2026

“GLM-5.3 is the most capable open-weights model for coding.”

Command Code

@CommandCodeAI

X Aug 14, 2026

“GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal... What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling.”

Jie Tang

@jietang · Founder and chief scientist, Zhipu AI

X Aug 19, 2026

“GLM 5.3 is finally here I had the chance to test it 24 hours before its official release, and what I noticed > it stays on track longer > handles complex coding tasks with fewer corrections > feels noticeably more reliable across multi step workflows..”

JUMPERZ

@jumperz

X Aug 14, 2026

“GLM 5.3 is actually a huge deal. They didn't make the underlying model bigger or fundamentally change the architecture. They just took GLM 5.2 and post trained it harder. Yet, that alone produced major gains, especially in complex coding, long-horizon agent tasks, and...”

AI Search

@aisearchio

X Aug 14, 2026

“GLM-5.3 is the most capable open-weights model for coding.”

Command Code

@CommandCodeAI

X Aug 14, 2026

Frequently asked questions

Is GLM-5.3 open source?

Not yet. Z.ai released GLM-5.3 on August 14, 2026 through the GLM Coding Plan and ZCode, and said the weights follow about two weeks after launch, once safety evaluation and hardening are complete. That points to roughly August 28, 2026. The Hugging Face listing reads Coming Soon. Z.ai has not announced a license for GLM-5.3, though GLM-5 and GLM-5.2 both shipped under MIT.

What is the difference between GLM-5.3 and GLM-5.2?

Only the post-training. Z.ai states that GLM-5.3 uses the same base model as GLM-5.2, so the 753 billion total parameters, roughly 40 billion activated per token, and the 1M token context window are unchanged. One extra month of scaling long-horizon RL environments produced the gains: Terminal Bench 3.0 from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, CyberGym from 77.2 to 84.5, and ExploitBench from 24.4 to 54.4.

How much does the GLM-5.3 API cost?

GLM-5.3 costs $1.40 per million input tokens, $0.26 per million cached input tokens, and $4.40 per million output tokens, the same rates as GLM-5.2. Through the GLM Coding Plan the model is billed on a points quota instead, with calls outside 14:00 to 18:00 UTC+8 on weekdays consuming half the standard points.

How does GLM-5.3 compare to Kimi K3, Claude Fable 5, and GPT-5.6 Sol?

It trades wins and is close on the headline coding test. Z.ai reports 88.2 on Terminal Bench 2.1 against 88.3 for Kimi K3, 88.0 for Fable 5, and 88.8 for GPT-5.6 Sol. GLM-5.3 leads CyberGym at 84.5 against 83.8 and 83.6, and tops GDPval-AA v2 at 1769. The closed models keep clear leads on the harder tests: Terminal Bench 3.0 is 28.3 against 33.7 for Fable 5 and 34.6 for GPT-5.6 Sol, and ExploitBench is 54.4 against 78.0 and 76.5.

Can I run GLM-5.3 on my own hardware?

Not until the weights are published, which Z.ai targeted for roughly August 28, 2026. Once they land, the base is identical to GLM-5.2, so the same single-node setups apply: 8x H200, 8x B200, or 8x MI355X in BF16 or FP8. Until then GLM-5.3 is reachable only through the Z.ai API, the GLM Coding Plan, and ZCode.

Keep exploring

From the blog

The Best Open Source LLMs in 2026