HomeAI models
Z.ai: GLM 4.7 Flash
Z.AI
•
GLM-4.7-Flash
•
Released 
January 2026

Z.AI
GLM-4.7-Flash

Open-weight 30B-A3B reasoning model with 200K context and 128K output for low-cost agentic coding and tool use.

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

Output Speed *

N/A
tok/s

Intelligence Index *

14.9
/ 100

Context Window *

202752
tokens

Input price

0.37
Anytoken

Output price

2.4
Anytoken
GLM-4.7-Flash: Agentic Coding and Reasoning at 3B Active Parameters GLM-4.7-Flash is Z.ai's lightweight member of the GLM-4.7 family: a 30B-A3B Mixture-of-Experts model with 31.2B total and about 3B active parameters, released with open weights under the MIT license. The reasoning configuration targets agentic coding, multi-step tool use and long-context text work, with a 200K-token context window and up to 128K output tokens. It is text-in, text-out. High-volume coding-agent and tool-calling workloads benefit most, provided they can tolerate thinking-mode latency. Access GLM-4.7-Flash via the AnyAPI.ai API

Performance

Agentic Coding Results Well Above Its 3B Active-Parameter Class

Z.ai reports 59.2 on SWE-bench Verified and 79.5 on τ²-Bench. On the same model card, Qwen3-30B-A3B-Thinking-2507 scores 22.0 and 49.0. These are provider-reported figures, so treat them as directional. The gap suggests the model was tuned for agent loops rather than pure chat. Only about 3B parameters are active per token, which keeps inference compute modest. In production it fits coding agents and tool-calling pipelines. It is a weaker fit for knowledge-heavy question answering.

Benchmarks

Independent Data: Index Score of 15, Strong Tool Use, Below-Median Speed

Artificial Analysis scores the reasoning variant 15 on its Intelligence Index, against 11 for the non-reasoning variant, so thinking mode materially lifts results. Reasoning also raises τ²-Bench Telecom to 98.8% and Terminal-Bench Hard to 22.0%, up from 3.8% without it. Measured output speed is around 71 tokens per second, below the 90 median for comparable models. Time to first token is about 1.5 seconds, but time to first answer token runs roughly 25–31 seconds because of reasoning.

Output Speed

*
N/A
tok/s

Intelligence Index

*
14.9
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
58
%

HLE *

Adherence to multi-step structured instructions.
8
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

GLM-4.7-Flash accepts text and returns text only. Z.ai documents a 200K-token context window and up to 128K output tokens. These are separate limits, and some third-party hosts cap them lower: Cloudflare lists 131,072 tokens of context. Thinking is enabled by default, which adds generation time before the final answer appears. Check the context and output caps of the endpoint you call before sizing long-document or long-generation workloads.
Verified Specifications — 
GLM-4.7-Flash
*
Input modalities
Text
output modalities
Text
Context window
202752
 tokens
Maximum output tokens
117964
Reasoning
Yes
Knowledge cutoff
January 2026
Pricing (standard)
0.37
 AnyTokens in
 / 
2.4
 AnyTokens out

Limitations & Trade-offs

Where GLM-4.7-Flash falls short

1
Reasoning latency before the answer. Artificial Analysis measures time to first token near 1.5 seconds, but time to first answer token of roughly 25–31 seconds, with output speed around 71 tokens per second, below the 90 median for comparable models. Interactive chat, voice and autocomplete suffer. Batch agents and background jobs tolerate it. For latency-sensitive paths, disable thinking where the endpoint allows it, or choose a non-reasoning or faster model.
2
Weak closed-book factual reliability. On Artificial Analysis, the reasoning variant records 16.2% AA-Omniscience accuracy and a 6.1% non-hallucination rate, which points to confident wrong answers when it lacks knowledge. A 3B-active design plausibly limits stored knowledge, though that is interpretation. Support bots and knowledge Q&A should ground answers with retrieval and verification. Prefer a larger model when unaided factual recall matters.
3
Text-only, mid-tier ceiling. The model takes text input only, so screenshots, charts, scanned documents and audio need a separate model. Its Intelligence Index score of 15 also sits well below GLM-4.7 at 22 and GLM-5 at 28 in Artificial Analysis comparisons. Hard multi-file engineering, research-grade reasoning and multimodal pipelines are better served by a larger or vision-capable model.
4
Limits differ by host. Z.ai documents 200K context and 128K output, but Cloudflare Workers AI lists 131,072 tokens of context and Venice lists 128K context with 16,384 output tokens. Some hosts also keep reasoning always on. Long-document and long-generation workloads can fail or truncate if you assume the official numbers. Confirm the limits and thinking behavior of your endpoint before production sizing.

Best-Fit Workloads

Where this model earns its place

01

Coding agents and issue triage

‍
Z.ai reports 59.2 on SWE-bench Verified, and Artificial Analysis measures 22.0% on Terminal-Bench Hard with reasoning, up from 3.8% without it. That suits patch-generation loops, repository bug triage and terminal-style agents where many steps run per task. The model's lower intelligence ceiling means complex multi-file refactors may need escalation to GLM-4.7 or a larger model.

02

Multi-step tool-calling pipelines

‍
The model supports function calling with automatic tool choice. Z.ai reports 79.5 on τ²-Bench, and Artificial Analysis measures 98.8% on τ²-Bench Telecom. Z.ai recommends Preserved Thinking for multi-turn agent tasks. This fits workflow agents that call internal APIs and reason across turns, as long as the reasoning delay is acceptable for background execution.

03

Long-context document analysis

‍
A 200K-token window lets the model read large codebases, logs or report sets in one request. Artificial Analysis AA-LCR scores 41.7% with reasoning versus 20.3% without, so thinking helps long-context work. Treat it as moderate rather than best-in-class, and ground answers against the source text, since factual recall without context is weak.

04

Open-weight prototyping and self-hosting paths

‍
The weights are MIT-licensed on Hugging Face. Lambda lists 62.5GB of weights and a single H100 or B200 as a minimal deployment, with more GPUs for the full 200K context. Teams can prototype through an API, then move to private infrastructure if data-control needs grow. Full-context self-hosting is memory-heavy.

Pricing in anytokens via AnyAPI
Input
0.37
₳
Output
2.4
₳
Cache write
—
₳
Cache read
0
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access GLM-4.7-Flash via AnyAPI.ai

Access GLM-4.7-Flash through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate GLM-4.7-Flash 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-Flash 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-Flash 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-Flash 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-Flash 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-Flash has a 200K-token context window, and Z.ai documents up to 128K output tokens. Context and output are separate limits. Some third-party hosts list lower caps, such as 131,072 tokens of context on Cloudflare Workers AI, so confirm the limits for the endpoint you use.

Yes. Artificial Analysis benchmarks it as GLM-4.7-Flash (Reasoning), scoring 15 on its Intelligence Index versus 11 for the non-reasoning variant. Thinking is on by default on the Z.ai API, and Z.ai recommends Preserved Thinking for multi-turn agent tasks. The trade-off is a long wait before the first answer token.

No. GLM-4.7-Flash accepts text input and produces text output only, according to Z.ai documentation and Artificial Analysis. For screenshots, charts or scanned documents, route those requests to a separate vision-capable model and use GLM-4.7-Flash for the text reasoning, coding or tool-calling steps.

GLM-4.7 is more capable, and GLM-4.7-Flash is faster. Artificial Analysis scores GLM-4.7 (Reasoning) at 22 versus 15 for Flash. Flash generates about 85 tokens per second against roughly 42 and reaches its first answer token in about half the time. Both list 200K context and 128K output.

Yes. The weights are on Hugging Face under the MIT license. It is a 30B-A3B Mixture-of-Experts model with about 31.2B total and 3B active parameters. Lambda lists 62.5GB of weights and a single H100 or B200 as a minimal deployment, while the full 200K context needs multiple GPUs.

* Benchmark data source: Artificial Analysis artificialanalysis.ai