Flagship agentic coding with 1M context
Run instantly. Pay only for what you use.
GLM 5.3 is Z.ai's flagship agentic coder, released as open weights under Z.ai's GLM-5.3 license: a 744B-parameter MoE with 40B active per token, a 1M-token context window, and up to 128K output tokens, with reasoning effort selectable per request — low, high, or max.
It is a pure post-training release on the same base model as GLM 5.2 — every gain comes from post-training — and the gains are long-horizon: SWE-Marathon climbs 19.4 → 42.5 and Terminal-Bench 3.0 jumps 4.6 → 28.3, so engineering runs that span hours hold together instead of unraveling.
At maximum reasoning effort it posts 88.2 on Terminal-Bench 2.1 — within 0.6 of GPT-5.6 Sol — 66.9 on DeepSWE 1.1, and the top published 84.5 on CyberGym, with strict function calling and JSON output in every mode.
88.2on Terminal-Bench 2.1
Within 0.6 of GPT-5.6 Sol on end-to-end terminal tasks — from open weights.
1Mtoken context window
With up to 128K output tokens — an entire repository in, whole modules out.
97%avg. cache hit rate on Entrim
Measured on coding-agent workloads — repeated context bills at the cache-read rate.
GLM 5.3 is strongest where engineering runs are measured in hours — marathon refactors, deep debugging sessions, and security analysis that has to hold a plan across thousands of steps.
88.2 on Terminal-Bench 2.1 and 66.9 on DeepSWE 1.1 — frontier-class terminal work from open weights.
SWE-Marathon climbs 19.4 → 42.5 over GLM 5.2 — multi-hour issue streams hold together to the end.
84.5 on CyberGym — the top published score, ahead of closed frontier models — for defensive audits at repo scale.
Low, high, and max efforts set thinking depth per request, defaulting to max — pay for depth only where a run needs it.
The hardest long-horizon engineering
42.5 on SWE-Marathon at max effort — more than double GLM 5.2 — for refactors and migrations that run for hours.
Long-running coding agents
Plans, edits, and verifies across long tool-call chains — 88.2 on Terminal-Bench 2.1 and 66.9 on DeepSWE 1.1.
Security review and vulnerability analysis
84.5 on CyberGym — the top published score, ahead of closed frontier models — for defensive code and dependency audits.
Whole-repo analysis and generation
The 1M-token window reads an entire repository, and 128K output tokens emit whole modules or reports in one response.
Latency-critical short replies
There is no non-thinking mode — even low reasoning effort spends tokens before the reply. Use DeepSeek V4.1 Flash in chat mode for fast conversational turns.
Tight output-token budgets
Reasoning traces run long — about 1.5× the median output tokens in Artificial Analysis testing. Use low reasoning effort for cost-sensitive replies.
Image, audio, or video inputs
GLM 5.3 is text-only. For agentic work grounded in screenshots or video, use GLM 5.3 Flash.
Turbo is Entrim's NVFP4 serving mode for GLM 5.3 — the same open weights, quantized only where it is safe, for higher throughput at a lower price.
Only the linear-operator weights and activations inside the MoE transformer blocks are stored in NVFP4 — attention, routing, embeddings, and everything else stays in higher precision, so 4-bit noise never touches the layers most sensitive to it.
That takes those weights from 8 bits per parameter to 4 — about a 1.66× smaller footprint on disk and in GPU memory, which is what makes the 1M-token context affordable to serve.
FP8 vs NVFP4, measured by NVIDIA on GLM 5.3
| Benchmark | FP8 | NVFP4 | Δ |
|---|---|---|---|
| GPQA Diamond | 91.79 | 92.68 | +0.89 |
| SciCode | 58.65 | 58.03 | −0.62 |
| LiveCodeBench V6 | 52.95 | 53.55 | +0.60 |
| AA-LCR (long context) | 69.75 | 69.63 | −0.12 |
| IFBench | 65.33 | 66.00 | +0.67 |
| Terminal Bench 2.1 | 85.67 | 86.80 | +1.13 |
Measured by NVIDIA on its GLM 5.3 NVFP4 checkpoint, served on SGLang. Entrim’s own evals will replace this table when they land.
The table above is NVIDIA's measurement, not Entrim's — validate outputs against your own evals before moving production traffic.
88.2Terminal-Bench 2.1
Agentic coding and shell tasks run end-to-end in a real terminal.
66.9DeepSWE 1.1
Multi-step software engineering fixes — up from 46.2 for GLM 5.2.
84.5CyberGym
Vulnerability analysis across 1,507 real codebases — the top published score.
Code examples
curl -X POST "https://api.entrim.ai/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "zai-org/GLM-5.3", "messages": [ { "role": "user", "content": "Why do they call it a building if it is already built?" } ] }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
Multimodal coding agents with 1M context
Agentic coding with image input and 1M context
Dense multimodal coding with thinking control