HomeAI models
Z.ai: GLM 4.7
Z.AI
•
GLM-4.7 (Reasoning)
•
Released 
December 2025

Z.AI
GLM-4.7 (Reasoning)

Open-weight MIT-licensed reasoning model for agentic coding, with 200K context and 128K output for long, tool-driven tasks.

Modality:
Text
PDF
model ID
z-ai/glm-4.7
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

0.00
tok/s

Intelligence Index *

22.2
/ 100

Context Window *

202752
tokens

Input price

3.6
Anytoken

Output price

13.2
Anytoken
GLM-4.7: Open-Weight Agentic Coding With 128K Output GLM-4.7 is Z.ai's December 2025 flagship text model, an MIT-licensed open-weight Mixture-of-Experts system with 358B parameters and interleaved thinking. It sits between GLM-4.6 and the later GLM-5 series. Its main draw is agentic coding: Z.ai reports 73.8% on SWE-bench Verified and 87.4 on τ²-Bench. It suits coding agents, terminal automation and multi-step tool workflows that need long outputs. Start building with GLM-4.7 on AnyAPI.ai

Performance

Agentic Coding Gains Over GLM-4.6, Per Z.ai's Own Scores

GLM-4.7's clearest strength is multi-step coding and tool use. Z.ai's model card reports 73.8% on SWE-bench Verified (GLM-4.6: 68.0%), 41.0% on Terminal Bench 2.0 (24.5%) and 87.4 on τ²-Bench (75.2). These are provider-reported results, so treat them as directional. The largest jumps are on terminal and tool-use tasks, which suggests better action sequencing in long agent runs. In production, that favors agent loops and repository-level fixes over single-shot chat.

Benchmarks

GLM-4.7 on Artificial Analysis: Above-Average Intelligence, Low First-Token Latency

Artificial Analysis (Intelligence Index v4.3.2) scores GLM-4.7 (Reasoning) at 22, an estimate. That is above the 18 median for similarly sized open-weight models, but well below Grok 4.5 (High) at 39. Median time to first token is 1.27s versus a 2.32s peer median. Reasoning pushes time to first answer token to roughly 21s in one snapshot. Provider choice matters: in January, AA measured Cerebras at 1,445 output tokens/s versus 430 for the fastest GPU provider, Fireworks. Index scores are not comparable across index versions.

Output Speed

*
0.00
tok/s

Intelligence Index

*
22.2
/ 100

MMLU *

Broad world knowledge and problem-solving
86
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
86
%

HLE *

Adherence to multi-step structured instructions.
27
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
89
%

Technical Specifications

What the model supports

GLM-4.7 takes text input and returns text only, with a 200K-token context window and up to 131,072 output tokens (about 128K). The output ceiling is the spec with the most production impact: full modules or long reports can fit in one response. Context limits vary by host. Some list 204,800 tokens, while NVIDIA's listing caps input and output at 131,072. Confirm limits on the route you use.
Verified Specifications — 
GLM-4.7 (Reasoning)
*
Input modalities
Text
PDF
output modalities
Text
Context window
202752
 tokens
Maximum output tokens
131072
Reasoning
Yes
Knowledge cutoff
December 2025
Pricing (standard)
3.6
 AnyTokens in
 / 
13.2
 AnyTokens out

Limitations & Trade-offs

Where GLM-4.7 (Reasoning) falls short

1
Text-only interface. GLM-4.7 accepts text input and produces text output. It has no image, audio or file modality on the model itself, so screenshot debugging, scanned-document extraction and UI-from-mockup flows need pre-processing or a separate vision model. This is no handicap for code and text agents, but multimodal products should choose a different model.
2
Reasoning latency before the answer. Artificial Analysis measures a fast median time to first token of 1.27s, but its time to first answer token includes reasoning: about 21s in one snapshot and over 46s in another. Interactive autocomplete, voice assistants and live chat suffer. Use a non-reasoning or lighter model for those, and keep GLM-4.7 for tasks where thinking pays off.
3
200K ceiling and host-dependent limits. The 200K window is smaller than the 1M-token window listed for Z.ai's newer GLM-5.2. Some hosts also list lower limits: NVIDIA's listing caps input and output at 131,072 tokens. Whole-monorepo ingestion or very long document sets may not fit, and long-context behavior was not independently verified here. Confirm the limit on your route.
4
Behind newer frontier models. Artificial Analysis rates Grok 4.5 (High) at 39 against GLM-4.7's estimated 22, and Z.ai's own table shows other models scoring up to 54.2 on Terminal Bench 2.0 versus GLM-4.7's 41.0. For the hardest agentic and reasoning tasks, a newer or proprietary model may be the better primary choice. GLM-4.7 fits better as a strong open-weight workhorse.

Best-Fit Workloads

Where this model earns its place

01

Coding agents and repository fixes

‍
Z.ai reports 73.8% on SWE-bench Verified and 66.7% on SWE-bench Multilingual, and the model reasons before each response and tool call. That suits agents that read issues, edit multiple files and run tests. These figures are provider-reported, and its Terminal Bench 2.0 score trails the best models, so keep a test harness in the loop.

02

Terminal and tool-driven automation

‍
GLM-4.7 scores 41.0% on Terminal Bench 2.0 (up from 24.5% for GLM-4.6) and 87.4 on τ²-Bench. The Z.ai API also streams tool-call arguments. Good fits are CLI coding agents, DevOps scripting and multi-step support or operations flows. Add validation and retry limits, since provider-reported benchmarks do not guarantee your tool schemas will behave the same.

03

Long single-pass generation

‍
With up to 131,072 output tokens, GLM-4.7 can produce full modules, long technical reports or large structured documents in one response. Few models offer this much output headroom. Reasoning adds latency, so run these jobs asynchronously rather than in live user-facing chat, and confirm the output limit on your hosting route.

04

Math and STEM reasoning

‍
Z.ai reports 95.7% on AIME 2025, 85.7% on GPQA-Diamond and 42.8 on HLE with tools. These are provider-reported results, but they indicate solid quantitative reasoning for tutoring, analytics assistants and scientific problem-solving pipelines. The model is text-only, so diagrams or charts must be described in text, and answers should be verified for high-stakes use.

Pricing in anytokens via AnyAPI
Input
3.6
₳
Output
13.2
₳
Cache write
—
₳
Cache read
0.66
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access GLM-4.7 (Reasoning) via AnyAPI.ai

Access GLM-4.7 (Reasoning) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate GLM-4.7 (Reasoning) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access GLM-4.7 (Reasoning) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test GLM-4.7 (Reasoning) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use GLM-4.7 (Reasoning) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use GLM-4.7 (Reasoning) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

GLM-4.7 has a 200K-token context window and supports up to 131,072 output tokens (about 128K). Some hosts list 204,800 tokens of context, and NVIDIA's listing caps input at 131,072, so confirm limits on your route. Context window and maximum output are separate limits.

No. GLM-4.7 accepts text input and returns text output. For images, use a separate vision-capable model. The model does support tool calling, streaming and reasoning output, but it has no native image, audio or video modality.

Z.ai reports 73.8% on SWE-bench Verified, 66.7% on SWE-bench Multilingual, 41.0% on Terminal Bench 2.0 and 84.9% on LiveCodeBench-v6. These are provider-reported results. GLM-4.6 scored 68.0% on SWE-bench Verified in the same table. Validate on your own repositories.

GLM-4.7 is open-weight under the MIT license, with weights published on Hugging Face. It is a 358B-parameter Mixture-of-Experts model, so self-hosting needs multi-GPU infrastructure. Supported inference frameworks include vLLM, SGLang and Transformers. Using the API avoids that hardware overhead.

On the Z.ai API, thinking is enabled by default through the thinking parameter, and the model applies Interleaved Thinking before responses and tool calls. Z.ai also describes Turn-level Thinking for adjusting depth per turn. Reasoning adds delay before the first answer token, so plan for that latency.

* Benchmark data source: Artificial Analysis artificialanalysis.ai