GLM-4.5-Air from Z.ai is the small sibling of GLM-4.5: 106B total parameters with just 12B active, against 355B total and 32B active for the full model. That shrink is what buys the $0.20 per Mtok input and $1.10 per Mtok output price, half the output cost of GLM-4.5 on the same 128K tokens (131,072) context. Inside Atlas, GLM-4.5-Air is the cheap slot: point small_model at it, keep a stronger model in the main slot, and the background traffic that Atlas generates constantly stops mattering to your bill.
What makes GLM-4.5-Air cheap enough for Atlas background traffic?
GLM-4.5-Air activates only 12B parameters per forward pass against 32B for GLM-4.5, and that architectural shrink is what buys the $0.20 per Mtok input rate. Z.ai charges one third of GLM-4.5's $0.60 while keeping the same 131,072 context and 98,304 output cap.
Atlas produces a steady stream of low-stakes model calls: chat titles, session summaries, subagent coordination. None of it needs a 355B model. GLM-4.5-Air is sized for exactly that work at $0.20 per Mtok input and $1.10 per Mtok output. The important detail is what Z.ai did not cut. GLM-4.5-Air keeps the full 131,072 token context and the 98,304 output cap of the paid GLM-4.5 tier, so it is not a toy-window model that chokes when Atlas hands a subagent a genuinely large retrieval result.
Can you self-host GLM-4.5-Air on a single node?
GLM-4.5-Air's 106B total parameter footprint is small enough that quantized self-hosting on a single high-VRAM node is actually plausible, unlike the 355B GLM-4.5. Z.ai ships GLM-4.5-Air under the MIT license, so the identical weights you call over the API can be run yourself.
The MIT license here means the cheap tier is not a crippled hosted-only SKU. Pull the MIT-licensed zai-org/GLM-4.5-Air weights, then register Ollama as a custom provider in Atlas with "limit": { "context": 131072, "output": 98304 } so Atlas budgets the window correctly. That declaration matters: Atlas indexes code by AST declarations using tree-sitter, not blind line windows, and it sizes retrieval against the configured context limit. Combined with Atlas building its code index using local Ollama embeddings, a self-hosted GLM-4.5-Air setup keeps code entirely off third-party servers.
Where does GLM-4.5-Air fall short of GLM-4.5?
GLM-4.5-Air's 12B active parameters versus GLM-4.5's 32B shows up on hard multi-step debugging, where GLM-4.5-Air drops steps that GLM-4.5 holds onto. The context window is identical at 131,072 tokens, so the gap is reasoning depth, not memory.
This is the honest limit of the Air tier, and it is worth understanding precisely. GLM-4.5-Air is not worse at reading code or at recalling what is in the window. It is worse at holding a long chain of dependent inferences: trace the null through three call sites, then reconcile it against the test that passes, then decide which of the two is wrong. Atlas drafts a plan in a read-only plan agent and asks before switching to a build agent, and that plan phase is exactly where 12B active parameters start to show. Put a stronger model in the main slot for that work.
Is GLM-4.7-FlashX a better cheap slot than GLM-4.5-Air?
In most cases yes. GLM-4.7-FlashX costs $0.07 per Mtok input and $0.40 per Mtok output with a larger 200,000 token context, which undercuts GLM-4.5-Air's $0.20 and $1.10 on both sides while giving Atlas more room to work.
GLM-4.7-FlashX is now cheaper still with a larger window, which makes GLM-4.5-Air a harder sell than it was. The case that survives is the open-weights case: GLM-4.5-Air is MIT licensed and self-hostable on a single high-VRAM node, so if your requirement is running the cheap slot on your own hardware rather than merely paying little for it, GLM-4.5-Air is the model that supports that. If you are happy calling a hosted API, price says take GLM-4.7-FlashX.
How do you wire GLM-4.5-Air into Atlas as the small model?
Set "small_model": "zai/glm-4.5-air" in atlas.json and keep a stronger model in "model". Atlas then routes titles, summaries, and subagent chatter to GLM-4.5-Air at $0.20 per Mtok input while the interactive loop stays on the heavier checkpoint.
The two-slot split is the whole point of GLM-4.5-Air. Atlas fans out work to subagents that can run in the foreground or in parallel background sessions, and that fan-out is what makes a cheap small_model economically significant rather than a rounding error. Export ZHIPU_API_KEY or run `atlas login` and choose Z.ai, then run `atlas models zai` and locate `glm-4.5-air` to confirm the id resolves before you edit config. Atlas lets you switch the active model and provider on the fly with favorites and recents, so the main slot stays flexible.
Setup
- 01Export ZHIPU_API_KEY, or run `atlas login` and choose Z.ai.
- 02Run `atlas models zai` and locate `glm-4.5-air` in the model list.
- 03Use GLM-4.5-Air as the cheap slot: set "small_model": "zai/glm-4.5-air" in atlas.json, with a stronger model in "model".
- 04For local use, pull the MIT-licensed `zai-org/GLM-4.5-Air` weights.
- 05Register Ollama as a custom provider with "limit": { "context": 131072, "output": 98304 } so Atlas budgets the window correctly.
Frequently asked questions
- how much does glm-4.5-air cost
- GLM-4.5-Air costs $0.20 per Mtok input and $1.10 per Mtok output, one third of GLM-4.5's input price and half its output cost.
- how many parameters does glm-4.5-air have
- GLM-4.5-Air is a 106B total parameter MoE with 12B active per forward pass, against 355B total and 32B active for GLM-4.5.
- can I run glm-4.5-air on one gpu node
- Quantized self-hosting of GLM-4.5-Air on a single high-VRAM node is plausible given its 106B total footprint, unlike the 355B GLM-4.5.
- what is the glm-4.5-air context window
- GLM-4.5-Air has a 128K tokens (131,072) context window and a 98,304 output cap, identical to the paid GLM-4.5 tier.
- how do I set glm-4.5-air as the small model in atlas
- Export ZHIPU_API_KEY, run `atlas models zai` to locate `glm-4.5-air`, then set "small_model": "zai/glm-4.5-air" in atlas.json with a stronger model in "model".
- glm-4.5-air vs glm-4.7-flashx
- GLM-4.7-FlashX is cheaper at $0.07 input and $0.40 output with a larger 200,000 context. GLM-4.5-Air's remaining edge is MIT-licensed self-hosting.
- is glm-4.5-air good at debugging
- GLM-4.5-Air struggles on hard multi-step debugging, where its 12B active parameters drop steps that GLM-4.5's 32B active holds onto.
Try Atlas in your terminal
The terminal-native AI coding agent. Free core, single binary.
Install AtlasRelated guides
Automate GitHub Issue and Pull Request Triage with Atlas (2026 Workflow)
How to automate GitHub issue and pull request triage with Atlas in 2026: the atlas github command checks the actor has admin or write permission before it does anything.
Atlas for Haskell in 2026
Atlas is a terminal-native AI coding agent for Haskell in 2026. Run it in a project with a .cabal file or a stack.yaml, let cabal build drive it, and review the diff.
Atlas vs Qodo: Choosing Your AI Coding Agent in 2026
Comparing Atlas, the terminal-native AI coding agent, with Qodo 2.0, the multi-agent PR reviewer, for developers in 2026. Evaluate features, pricing, and workflow.
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 vs Blackbox AI: Choosing Your AI Coding Agent in 2026
Comparing Atlas, the terminal-native AI coding agent, with Blackbox AI, a VS Code agent with 4.7 million installs, for developers in 2026. Evaluate features, pricing, and workflow.
Atlas vs Poolside: Terminal AI Coding Agents in 2026
Compare Atlas and Poolside, two terminal AI coding agents in 2026. Atlas offers a free core and robust safety features, while Poolside provides open-weight Laguna models and on-prem deployment options.
Debug a Single Failing Test with Atlas in 2026
How to debug one failing test with Atlas in 2026: run it in isolation with bash, walk the call graph with the lsp tool, and fix the code, not the assertion.
Extract a Shared Helper from Duplicated Code with Atlas (2026 Workflow)
How to extract a shared helper from duplicated code with Atlas in 2026: codebase_search finds the copies by meaning, write creates the module, apply_patch swaps each call.