Model card
Kodwen 1 (beta)Beta
Kodwen 1 is Kodwen’s own coding model for agentic software-engineering work, in beta.
| API name | Kodwen 1 (beta) (moving alias) |
|---|---|
| Status | Opt-in beta, not the default behind Kodwen |
| Built for | Agentic software engineering: reading a repo, editing code, running tests |
| Context | Up to 1M tokens |
| Input and tools | Text; tool calling |
| Weights | Not released; available only through the Kodwen Api |
What it is
Kodwen 1 is a coding model for agentic software-engineering work: reading a repository, editing code and running tests. It is in beta and its behaviour can change between snapshots.
Kodwen 1 (beta) is opt-in: you choose it by name in the API, the desktop app, the VS Code extension or any compatible client. Kodwen keeps its current default models, and Kodwen 1 is not part of any Kodwen name.
Context and tools
Text and tool calling, with a context window of up to 1M tokens — long enough to hold a repository and its test output in one request.
Image inputs are handled by an approved companion; the final answer and the tool calls come from Kodwen 1.
Scope
- Agentic software engineering: repository reading, multi-file edits, running the project’s own tests.
- Strongest on Python repositories today; nothing is claimed for other languages or domains.
- Available in the API, the desktop app, the VS Code extension or any OpenAI-compatible client.
- Runs behind the same gateway guardrails, AI-generated markers and usage records as every Kodwen model.
- The service is provided under the Kodwen Terms of Service.
Evaluation
Two internal evaluations, reported as in the model card. They are a regression guard, not a capability claim. No version of Kodwen 1 is better than the one before it beyond measurement noise, and none is worse. No comparison with any other model is made.
| Model | Thinking off | Thinking on | Δ off (95 % CI) | Δ on (95 % CI) | Verdict |
|---|---|---|---|---|---|
| Base | 74.9 % | 86.8 % | — | — | Reference |
| Kodwen 1 (beta)served as Kodwen 1 (beta) | 75.2 % | 88.2 % | +0.3 (−1.4, 1.9) | +1.4 (−0.1, 3.0) | QUALITY-NEUTRAL |
| Kodwen 1 (alt. snapshot) | 74.6 % | 87.0 % | −0.3 (−1.9, 1.4) | +0.3 (−1.3, 2.0) | QUALITY-NEUTRAL |
| Kodwen 1 (alt. snapshot 2) | 75.8 % | 87.3 % | +0.9 (−0.6, 2.3) | +0.6 (−1.0, 2.1) | QUALITY-NEUTRAL |
| Slice | Tasks (n) | Base | Kodwen 1 (beta) | Δ (95 % CI) | Verdict |
|---|---|---|---|---|---|
| Primary | 50 | 1.8 % | 1.7 % | −0.2 (−1.8, 1.5) | HOLD |
| Repo-disjoint from training | 43 | 1.4 % | 1.7 % | +0.4 (−1.2, 1.9) | HOLD |
- holdout-expanded-v1 is contaminated and saturated, so it is used only as a regression guard.
- Every model scores 2 % or less on the hard pool, so it supports no capability claim either way. Verdict: HOLD.
- Harness errors: 3 of 702 rows (0.4 %) per arm; two runs that crashed were excluded.
Where and how it runs
- Kodwen 1 runs in Kodwen’s own cloud; prompts and outputs are not sent to any model publisher.
- Capacity: about 8 concurrent streams per replica before throughput drops.
- No failover: Kodwen 1 (beta) has a single backend. If the replica is down, the request fails rather than being sent to another model.
- The same gateway guardrails, AI-generated markers and usage records apply as for Kodwen.
Limitations
- Beta: behaviour may change between snapshots, and the Kodwen 1 (beta) alias moves.
- No measured gain over earlier snapshots on any evaluation above.
- Quality is sensitive to the engine and the thinking mode.
- It can produce wrong, insecure or non-compiling code, and may reproduce public code seen during its pre-training. Review its output.
- Romanian and other non-English quality has not been evaluated.
- Not designed or assessed for high-risk uses or as the sole basis for decisions about people.
Version
- Model card 0.2, 24 October 2026.
- Behind the moving alias Kodwen 1 (beta); dated snapshots will be named Kodwen 1-YYYY-MM-DD.
- Weights are not released.
- Provided under the Kodwen Terms of Service.