GLM-4.5 from Z.ai was the July 2025 flagship that put MIT-licensed Chinese open weights on the coding map, and in 2026 it still runs Atlas at $0.60 per Mtok input and $2.20 per Mtok output. GLM-4.5 is a 355B total parameter MoE with 32B active, a 128K tokens (131,072) context, and a 98,304 output cap. Inside Atlas, GLM-4.5 is the pinned, reproducible open-weights baseline: the MIT license means the model you benchmark today is one you can redistribute, fine-tune privately, and keep running on your own hardware indefinitely.
What does the MIT license on GLM-4.5 let you do?
The MIT license on GLM-4.5 permits commercial use, redistribution, and private fine-tuning with no vendor lock. Z.ai shipped the full 355B total / 32B active MoE under that license in July 2025, which is why GLM-4.5 became the reference open-weights coding model for teams that cannot depend on a hosted-only SKU.
License terms are a real engineering constraint, not a legal footnote. A team that fine-tunes GLM-4.5 on its own codebase owns that derivative outright under MIT. A team that ships GLM-4.5 inside a product can redistribute it. And a team whose provider raises prices or retires an endpoint can keep serving the identical weights. Pair that with Atlas, which can build its code index with local Ollama embeddings, keeping code off third-party servers, and permission-gates every tool call against allow, ask, and deny rules before it runs, and you have a stack where nothing about your code has to leave your infrastructure.
How big is the GLM-4.5 output cap?
GLM-4.5 has a 98,304 max output token cap, which was unusually high for a 2025-era model and is enough for whole-file rewrites in a single pass. The 128K tokens (131,072) context window feeds it.
Output ceilings decide whether an agent can finish a job in one turn. GLM-4.5's 98,304 tokens of output is generous relative to its 131,072 context, so a whole-file rewrite lands in one pass rather than being stitched across turns. Atlas computes a unified diff for every file edit and surfaces it for approval before writing, so a large GLM-4.5 output arrives as a reviewable diff rather than a blind overwrite, and Atlas snapshots file changes as git patches so edits can be diffed and rolled back if the rewrite went the wrong direction.
How much does GLM-4.5 cost compared to GLM-4.7?
GLM-4.5 costs $0.60 per Mtok input and $2.20 per Mtok output, and GLM-4.7 charges the identical $0.60 per Mtok input. Z.ai set the family's price floor with GLM-4.5 in July 2025 and has held that input rate through later releases.
GLM-4.5 is where the GLM price floor was established, and it has proven durable. Reasoning enabled at $0.60 per Mtok input was aggressive in 2025 and remains competitive in 2026. The uncomfortable consequence for GLM-4.5 specifically is that its own family caught up without charging more: GLM-4.6 offers a larger window and better coding scores at the identical $0.60 / $2.20 price. If you are choosing on price alone, there is no premium to pay for moving up the line, which is the strongest argument against staying on GLM-4.5.
Is the GLM-4.5 context window too small in 2026?
GLM-4.5's 131,072 token context is now the smallest in the live GLM lineup. Both GLM-4.6 and GLM-4.7 moved to 204,800 tokens, so GLM-4.5 gives Atlas roughly 64 percent of the working set that its successors do.
Window size matters in Atlas because Atlas indexes code by AST declarations using tree-sitter, not blind line windows, and searches code with hybrid semantic and keyword retrieval fused by reciprocal rank fusion. Those retrieved declarations compete for space. On a 131,072 token GLM-4.5 context you fit fewer of them, which means either tighter retrieval or more turns. For a small service this is a non-issue. For a large repository where Atlas is pulling in dozens of declarations plus their tests, the 204,800 window of GLM-4.6 and GLM-4.7 is a material advantage at the same price.
When should you still pick GLM-4.5?
Pick GLM-4.5 when you need a pinned, reproducible checkpoint. GLM-4.6 beats it on window and coding scores at the identical $0.60 / $2.20 price, so the remaining case for GLM-4.5 in 2026 is reproducibility, an existing fine-tune, or a deployment already validated on these exact MIT weights.
Reproducibility is a legitimate reason to stay put. If you have benchmarked a workflow, tuned prompts against GLM-4.5's behavior, or shipped a fine-tune derived from zai-org/GLM-4.5, then moving generations means revalidating all of it. GLM-4.5 stays available and stays MIT, which is exactly the property that makes pinning viable. Self-hosting the weights is a serious undertaking though: at 355B total parameters, GLM-4.5 is a multi-GPU deployment, not a laptop one. Teams that want open weights on modest hardware should look at the smaller GLM-4.5-Air instead.
Setup
- 01Get a Z.ai key, then export ZHIPU_API_KEY or run `atlas login` and select Z.ai.
- 02Run `atlas models zai` to confirm that `glm-4.5` resolves from the registry.
- 03Set "model": "zai/glm-4.5" in atlas.json.
- 04To run the MIT weights on your own hardware instead, pull `zai-org/GLM-4.5` from HuggingFace.
- 05Plan for a multi-GPU deployment if you self-host, since GLM-4.5 is 355B total parameters and is not a laptop model.
Frequently asked questions
- how much does glm-4.5 cost per million tokens
- GLM-4.5 costs $0.60 per Mtok input and $2.20 per Mtok output on Z.ai, the same input rate that the much later GLM-4.7 charges.
- is glm-4.5 open source and can I use it commercially
- Yes. GLM-4.5 is released under the MIT license, which permits commercial use, redistribution, and private fine-tuning with no vendor lock.
- what is the glm-4.5 context window
- GLM-4.5 has a 128K tokens (131,072) context window and a 98,304 max output cap. It is the smallest window in the live GLM lineup.
- how do I use glm-4.5 with atlas
- Export ZHIPU_API_KEY or run `atlas login` and select Z.ai, confirm the model with `atlas models zai`, then set "model": "zai/glm-4.5" in atlas.json.
- glm-4.5 vs glm-4.6 which should I use
- GLM-4.6 offers a larger 204,800 window and better coding scores at the identical $0.60 / $2.20 price, so GLM-4.5 mainly survives as a pinned, reproducible checkpoint.
- can I self host glm-4.5
- Yes. Pull `zai-org/GLM-4.5` from HuggingFace. At 355B total parameters it is a multi-GPU deployment, not something you run on a laptop.
- how many parameters does glm-4.5 have
- GLM-4.5 is a Mixture of Experts model with 355B total parameters and 32B active per forward pass.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Atlas vs Sourcery: Terminal AI Coding Agents in 2026
Compare Atlas and Sourcery for 2026. Atlas offers terminal-native AI coding with permission-gated tools and local embeddings. Sourcery provides Python-first LLM review and rules-based refactoring, starting at
Atlas for WebAssembly: Terminal-Native AI Coding in 2026
Atlas is a terminal-native AI coding agent for WebAssembly. Audit wasm-bindgen exports, batch calls across the JS boundary, and run wasm-pack test --node in 2026.
Atlas for Electron: Terminal-Native AI Coding for Main, Preload, and Renderer in 2026
Atlas is a terminal-native AI coding agent for Electron in 2026, where the main and renderer split, contextIsolation, and preload bridges are the security model.
Atlas vs Mistral Vibe for Code: Terminal AI Coding Agents in 2026
Compare Atlas and Mistral Vibe for Code in 2026. Atlas offers terminal-native TUI, explicit diffs, and BYO models. Mistral Vibe for Code provides a four-model stack, multi-platform access, and EU data sovereignty.
Atlas for Java in 2026
Adopt Atlas, the terminal-native AI coding agent, for Java development in 2026. Enhance your workflow with intelligent code search, refactoring, and robust safety features for Maven and Gradle projects.
Atlas for Groovy: A Terminal-Native AI Coding Agent for Gradle, Spock, and Jenkins in 2026
Atlas is a terminal-native AI coding agent for Groovy in 2026. It reads build.gradle closures and Jenkinsfiles, writes Spock specs, runs ./gradlew test, and applies Spotless.
Atlas vs OpenAI Codex CLI: Terminal AI Coding Agents in 2026
Comparing Atlas and OpenAI Codex CLI for terminal AI coding in 2026. Atlas offers a TUI, permission-gated tools, and BYO models, while OpenAI Codex CLI provides sandboxed execution and OpenAI model integration.
Atlas for Clojure: A Terminal-Native AI Coding Agent for deps.edn and Kaocha in 2026
Atlas is a terminal-native AI coding agent for Clojure in 2026. It reads deps.edn aliases and namespace requires, runs clojure -M:test with Kaocha, and applies cljfmt.