⚡ SIMPLETI AI LABS · OFFICIAL RELEASE OCTOBER 2026
A Qwen3.8-27B fine-tune that answers code-change requests with SEARCH/REPLACE patches instead of whole files. Research release: so far only one live BF16 run of 120 tasks has been measured.
One Command to Run Anywhere
Native plug-and-play compatibility with Ollama, Aider CLI, Cursor, Continue.dev, and OpenCode.
Measured results
These are the only results measured so far: one live run of the full BF16 model on 120 tasks, on an NVIDIA RTX PRO 6000 Blackwell GPU. Earlier figures that did not come from a live run have been withdrawn.
| Category | Tasks | Unit tests passed | Patches applied |
|---|---|---|---|
| Surgical diff and AST precision | 30 | 17 / 30 | 20 / 30 |
| Functional correctness and edge cases | 30 | 6 / 30 | 27 / 30 |
| Ghost APIs and deprecated-code traps | 30 | 27 / 30 | 30 / 30 |
| Adversarial and out-of-distribution | 30 | 6 / 30 | 27 / 30 |
| Total | 120 | 56 / 120 | 104 / 120 |
Not evaluated: SWE-bench, Aider, comparison with other models or with the base model, and the Q4_K_M build that Ollama serves. The results above are for the BF16 weights.
Proprietary Architecture & Engineering Discipline
Engineering built to stop whole-file hallucination and keep maximum precision on every patch.
Answers with SEARCH/REPLACE blocks that change only the lines that need to change, instead of rewriting the whole file.
On the live BF16 run the output averaged 69.9 tokens per task (maximum 103).
Research release: a small, templated training set, no comparison with the base model yet, and no public benchmarks run.
Tuned out-of-the-box for Aider, Cursor, Continue.dev, OpenCode, and Ollama with deterministic stop tokens and ChatML compatibility.