GLM
Listed by the current MonkeyCode project README. Verify the specific model identifier and provider route available to your installation.
MODEL GUIDE · VERIFIED 2026-08-05
MonkeyCode supports model choice at the platform layer.
The public project currently names GLM, Kimi, MiniMax, Qwen, and DeepSeek. Exact model versions, endpoints, pricing, regions, limits, and availability can change and must be checked in the deployment you evaluate.
Documented model families
A current README list—not a permanent compatibility guarantee.
The names below reflect the public project source checked on 2026-08-05. “Supported” should not be interpreted as every version, region, account type, or deployment mode.
Kimi
Listed by the current MonkeyCode project README. Verify the specific model identifier and provider route available to your installation.
MiniMax
Listed by the current MonkeyCode project README. Verify the specific model identifier and provider route available to your installation.
Qwen
Listed by the current MonkeyCode project README. Verify the specific model identifier and provider route available to your installation.
DeepSeek
Listed by the current MonkeyCode project README. Verify the specific model identifier and provider route available to your installation.
Primary source: MonkeyCode project README ↗
Selection matrix
Verify the route, not only the model name.
Two deployments can select the same model family and still have different privacy, reliability, latency, and cost characteristics.
| Decision | Question to answer | Evidence to retain |
|---|---|---|
| Availability | Which exact model IDs and versions are selectable in this release and region? | Configuration screenshot and successful test request. |
| Routing | Does traffic go directly to a provider, through a private gateway, or to a locally operated endpoint? | DNS, proxy, firewall, and request-path evidence. |
| Credentials | Who owns the account and how are keys scoped, stored, rotated, and revoked? | Secret inventory and rotation test. |
| Data handling | What source context is sent and what are the provider's retention and training terms? | Current contract, policy, and request sample. |
| Operations | What are the rate limits, fallback behavior, observability, and outage procedure? | Load test, failure test, and runbook. |
| Cost and quality | Which model meets acceptance criteria at an acceptable total task cost? | Same-task pilot results, including retries and review. |
Model evaluation
Test with the same repositories and acceptance criteria.
A model leaderboard does not measure your environment, tools, policies, or reviewer effort.
- STEP 1Select bounded tasks
Use representative bug fixes, tests, documentation, and small features with written acceptance criteria.
- STEP 2Hold the workflow constant
Use the same repository state, environment, tools, prompts, and retry policy.
- STEP 3Measure accepted outcomes
Record pass rate, human corrections, latency, tokens, model cost, compute, and review time.
- STEP 4Test failure behavior
Exercise quota errors, timeouts, unavailable models, credential revocation, and fallback routes.
Common questions
Models and routing, answered.
Source-backed direct answers: can it use local models?, can it run offline?
Which AI models does MonkeyCode support?
The public project README currently lists GLM, Kimi, MiniMax, Qwen, and DeepSeek. Exact model identifiers, versions, regions, limits, and availability can change, so confirm what is selectable in the release you evaluate.
Can MonkeyCode use local or private models?
The self-hosting guide describes the model path as an external API, a private gateway, or a locally operated model, so private or in-network models are a documented option. Confirm which model IDs and endpoints your release actually supports.
Can MonkeyCode run models fully offline?
The project documents offline deployment, which requires the model to run on an in-network endpoint rather than an external API. Verify the offline model path against your specific release and hardware.
Does the model choice change data privacy?
Yes. Two deployments can select the same model family and still differ in routing, retention, latency, and cost. Map where inference traffic goes and what the provider retains before trusting a route with sensitive code.
How should I compare models for my codebase?
Run the same bounded tasks with an identical workflow and acceptance criteria, and measure accepted outcomes, human corrections, latency, token and compute cost, and failure behavior — not a public leaderboard.
MODEL FACTS CHANGE