What Is GLM 4.7? Specs, Pricing and Where It Fits
Nov 19, 2026

What Is GLM 4.7? Specs, Pricing and Where It Fits

What is GLM 4.7? Z.ai's December 2025 coding model: 200K context, MIT weights, Flash and FlashX variants, real API prices, and when to pick it over GLM-5.

What is GLM 4.7? Specs, variants and pricing of Z.ai's coding model

You open a project config and find model: glm-4.7. Or a tool's model picker offers GLM-4.7 next to GLM-5.3 at a fraction of the price. Or a free tier promises "GLM-4.7-Flash, completely free" and you wonder what the catch is.

What is GLM 4.7? It's a coding-focused large language model from Z.ai (Zhipu AI), released in December 2025. It has open weights under the MIT license, a 200K-token context window, and three variants: the full GLM-4.7, a cheaper FlashX, and a free Flash. It's the last major release before the GLM-5 generation, and it's still listed and priced on Z.ai's API.

This article covers what GLM 4.7 is, what changed from GLM-4.6, how the three variants differ, what it costs, and the question most explainers skip: whether it still makes sense now that GLM-5.x exists. It's based on Z.ai's official GLM-4.7 model cards, developer documentation, pricing page and Coding Plan documentation, all reviewed November 19, 2026. Benchmark figures are Z.ai's own published results. We did not re-run them. GLM5 publishes this article and runs independently of Z.ai.

GLM 4.7 at a Glance

GLM-4.7
Made byZ.ai (Zhipu AI)
ReleasedDecember 2025 (official repository published December 22, 2025)
TypeOpen-weight text LLM, text in / text out
Context window200K tokens
Max output128K tokens
ParametersAbout 358B in the published checkpoint
LicenseMIT
Built forAgentic coding, tool use, terminal tasks, UI generation
VariantsGLM-4.7, GLM-4.7-FlashX, GLM-4.7-Flash

The parameter figure comes from the official Hugging Face repository's file metadata. Z.ai's GLM-4.7 model card points to the GLM-4.5 technical report rather than restating the architecture. Treat any other parameter count you see as unverified.

What Is GLM 4.7 Built For?

Z.ai's own pitch for GLM-4.7 is "your new coding partner," and the official model card backs that with coding gains over GLM-4.6:

Benchmark (Z.ai's published results)GLM-4.6GLM-4.7Change
SWE-bench Verified68.073.8+5.8
SWE-bench Multilingual53.866.7+12.9
Terminal Bench 2.024.541.0+16.5
HLE (with tools)30.442.8+12.4
τ²-Bench (tool use)75.287.4+12.2

In plain terms, the biggest jumps are in terminal work (running commands in a real shell) and multilingual coding. Those are the areas where coding agents tended to fail. Z.ai also highlights cleaner webpage and slide layouts ("vibe coding") and better results inside agent frameworks such as Claude Code, Kilo Code, Cline and Roo Code.

What it isn't: GLM-4.7 is text-only. Z.ai's documentation lists text as the only input and output modality. For image input, Z.ai sells separate vision models.

The Feature That Matters Most: Preserved Thinking

GLM-4.7's most useful change for developers isn't a benchmark score. It's how the model handles reasoning across a conversation. The model card describes three modes:

  • Interleaved Thinking: the model reasons before every response and every tool call. This was introduced in GLM-4.5 and strengthened here.
  • Preserved Thinking: in coding-agent sessions, the model keeps its earlier reasoning blocks across turns and reuses them instead of re-deriving everything from scratch.
  • Turn-level Thinking: you can switch reasoning on or off per turn, so quick requests stay fast and cheap while hard ones get full reasoning.

Why Preserved Thinking matters: in a long agent session, a model that rebuilds its reasoning every turn can drift. It forgets why it chose an approach, or contradicts a decision from ten steps back. Keeping the reasoning reduces that drift. There is one practical trap. If your agent framework or proxy strips reasoning content from the message history between turns, Preserved Thinking has nothing to preserve. Check that your client passes it back.

GLM-4.7 vs GLM-4.7-FlashX vs GLM-4.7-Flash

All three share the 200K context and 128K output limits in Z.ai's documentation. They differ in size, speed and price:

VariantWhat it isZ.ai API price (per 1M tokens)Best for
GLM-4.7Full-size model$0.60 input / $0.11 cached / $2.20 outputHarder coding and agent tasks
GLM-4.7-FlashXLightweight, high-speed version$0.07 input / $0.01 cached / $0.40 outputHigh-volume, latency-sensitive calls
GLM-4.7-Flash30B-total, 3B-active MoE modelFreePrototyping, light tasks, local deployment

What is GLM-4.7-Flash, exactly? Its model card describes a 30B-A3B mixture-of-experts model: about 30 billion parameters in total, with roughly 3 billion active per token. Z.ai calls it the strongest model in the 30B class. It's small enough to run on far more modest hardware than the full model, and it's free on Z.ai's API. Expect it to trail the full GLM-4.7 on difficult, multi-step work.

Rule of thumb: use Flash to find out whether a task is easy. If Flash handles it, stay there. Escalate to FlashX for throughput or to the full model for difficulty only when Flash's answers start failing your checks.

Is GLM 4.7 Free?

Partly. Here's what's free and what isn't:

  • The weights are free to download under the MIT license. But the full model's roughly 358B-parameter checkpoint needs multi-GPU hardware to serve.
  • GLM-4.7-Flash is free on Z.ai's API, per the pricing page.
  • GLM-4.7 and GLM-4.7-FlashX are paid per token on Z.ai's API, at the rates above.

Is GLM 4.7 Open Source?

Yes, in the open-weight sense. Both GLM-4.7 and GLM-4.7-Flash publish their weights on Hugging Face under the MIT license, which permits commercial use under its terms. Z.ai also publishes deployment code in its GLM-4.5 GitHub repository, which covers the 4.5 through 4.7 releases.

How to Run GLM 4.7 Locally

The model card supports vLLM, SGLang and Transformers for local deployment. When GLM-4.7 launched, vLLM and SGLang needed their main-branch (nightly) builds to recognize it. The official vLLM example serves the FP8 checkpoint across four GPUs with tensor parallelism.

Realistic options by hardware:

  • Multi-GPU server: the full GLM-4.7, ideally the FP8 checkpoint, with vLLM or SGLang.
  • Single workstation GPU: GLM-4.7-Flash is the practical choice.
  • No GPU budget: use a hosted API. GLM-4.7-Flash is free on Z.ai, and the GLM5 API serves glm-4.7 as well.

Where GLM 4.7 Fits Now That GLM-5 Exists

GLM-4.7 was the top GLM model for less than two months. GLM-5 followed in February 2026 and, per its own model card, roughly doubled the scale of the GLM-4.5 generation: 744B total parameters with 40B active, up from 355B with 32B active. GLM-5.2 then added a 1M-token context, and GLM-5.3 is now Z.ai's flagship.

So should you still use GLM 4.7? It depends on what you're doing:

SituationRecommendation
Existing app running well on GLM-4.7Keep it, and test GLM-5.x on a sample of real traffic before switching
Cost-sensitive coding or agent workOn Z.ai, GLM-4.7 costs about half as much per token as GLM-5.2, or less. Worth testing
You need more than 200K tokens of contextMove to GLM-5.2 or GLM-5.3
Free or near-free experimentationGLM-4.7-Flash, or the newer GLM-5.3-Flash
Hardest long-horizon agent tasksUse the GLM-5 generation

One routing detail to know. On Z.ai's GLM Coding Plan, requests for GLM-4.7 are automatically routed to GLM-5.3-Flash, per the Coding Plan overview. If you configure a coding tool with a Coding Plan and select glm-4.7, you're actually getting GLM-5.3-Flash. To use GLM-4.7 itself, call it through a pay-as-you-go API.

To see how the generations compare in detail, read GLM 5.2 vs GLM-4. For the family as a whole, see what GLM is.

Try GLM 4.7 Against a Newer GLM

The fastest way to decide between GLM-4.7 and the GLM-5 line is to give both the same task. GLM5's API serves glm-4.7 alongside glm-5.2 and glm-5.3. Same base URL and same request format, so switching is a one-word change in your code. Get the model IDs from the GLM5 API docs. If you'd rather compare in a browser first, open GLM5 Chat and run your test prompt against a current GLM model as a baseline.

Frequently Asked Questions

Who made GLM 4.7?

Z.ai, the international brand of Zhipu AI. It's part of the same GLM family as GLM-5.2 and GLM-5.3.

When did GLM 4.7 come out?

December 2025. The official Hugging Face repository was published on December 22, 2025.

How many parameters does GLM 4.7 have?

The published checkpoint contains about 358 billion parameters, according to Hugging Face's file metadata. GLM-4.7-Flash is a separate 30B-total, 3B-active model.

What is GLM 4.7's context window?

200K tokens, with up to 128K output tokens, for all three variants, per Z.ai's documentation.

Is GLM 4.7 good?

For coding, it was a major step up from GLM-4.6. Z.ai reports 73.8 on SWE-bench Verified and 41.0 on Terminal Bench 2.0. Newer GLM-5.x models are stronger on long-horizon work, but GLM-4.7 remains a cost-effective option for many coding tasks.

Is GLM 4.7 multimodal?

No. It's text in, text out. Use one of Z.ai's vision models for image input.

The Bottom Line

GLM 4.7 is Z.ai's December 2025 coding model. It brought open MIT weights, a 200K context, big gains in terminal and multilingual coding, and Preserved Thinking for steadier agent sessions. It has since been overtaken by the GLM-5 generation, but its lower price and free Flash variant keep it relevant for cost-sensitive work.

If you're deciding whether to keep it, test rather than guess: run the same prompt on GLM-4.7 and GLM-5.x through the GLM5 API and let your own results decide.

Sources

Reviewed November 19, 2026. Prices, availability and routing rules can change. Confirm them on Z.ai's official pages before you commit. Benchmark numbers are Z.ai's published figures, not independent tests.

Start Using GLM 5 Today

Try GLM 5 free — reasoning, coding, agents, and image generation in one platform.