OpenMates Docs Open Chat

Cerebras Qwen 3.8 27B Evaluation

Cerebras Qwen 3.8 27B Evaluation Catalog OpenMates exposes as a Cerebras-primary, low-capability, explicit-selection-only model. The alias remains mapped to ...

[T:documentation.sender_name]

Cerebras Qwen 3.8 27B Evaluation

Catalog

OpenMates exposes alibaba/qwen-3.8-27b as a Cerebras-primary, low-capability, explicit-selection-only ai.ask model. The @fast alias remains mapped to alibaba/qwen3-235b-a22b-2507 until a stronger comparison supports changing it.

Official Cerebras documentation lists qwen-3.8-27b at approximately 1,500 tokens/second, with 128k paid context, 40k paid maximum output, and developer pricing of $0.99/M input and $1.49/M output tokens. It supports streaming, tool calling, structured output, reasoning controls, and image input.

Sources checked 2026-09-03:

Real Product-Path Benchmark

The benchmark used the deployed OpenMates CLI, the quick-code real-life TypeScript task, and --parallel 1 to avoid cross-model attribution problems. The OpenMates benchmark judge failed for the candidate, so the deterministic task result and observed duration are the reliable quality signals here.

Model Run Duration Deterministic result Judge
Cerebras Qwen 3.8 27B 548ac951-47ea-4b6d-a121-a3931e622923 23.6s Failed: returned an embed reference instead of code unavailable
Current @fast Qwen 3 256B be7ac101-2cf9-458d-a541-8303eb3efa25 19.1s Passed 5/5

The earlier two-model parallel run (c6088de1-ed43-46b0-92a2-9d9f340ab839) is retained as an operational finding only: model output attribution became ambiguous, so it is not used for the decision.

Recommendation

Do not replace @fast yet. In this product-path sample, Qwen 3.8 27B was approximately 23% slower end-to-end and failed the requested code-output contract, while the current Qwen 3 256B passed and was faster. Qwen 3.8 27B is available for explicit testing and should be re-evaluated with a larger, sequential suite after the benchmark judge and embed-output handling are stable.