Peculiar-Ragdoll's Tiel Ties Anthropic's Opus on Code Fixes Locally
A 22 GB quantized MoE coder matches Opus 4.6 medium on SWE-bench-Live while fitting in a single 24 GB consumer GPU.
Takeaways
- Tiel-Coder-35B-A3B is a re-quantized Ornith-1.5 MoE that solves 12/25 SWE-bench-Live problems, matching Opus 4.6 medium.
- Runs locally at 22 GB in 4-bit on a single 24 GB consumer GPU.
- Scores 67.2 on Claw-Eval multi-turn, beating its base Ornith-1.5 (65.3) and Nail (60.5).
- Trades knowledge for brevity: 73.7 on MMLU-Pro vs Nail's 84.0 at same quantization.
- Inherits Ornith's vision tower via an unmodified 903 MB BF16 projector for screenshots.
- A separate MTP variant adds a trained multi-token-prediction head for speculative decoding runtimes.
Peculiar-ragdoll, a small quantization lab, has released the Tiel model repo, a repackaged version of the Ornith-1.5-35B-A3B mixture-of-experts model. Tiel keeps Ornith’s weights while adding coding-focused quantization and a compact chat template. On the maintainers’ 25-issue SWE-bench-Live evaluation, the 4-bit, 22.4 GB build solved 12 issues, tying Anthropic’s Opus 4.6 medium configuration and running on a single 24 GB GPU.
The Hugging Face repository has recorded nearly 400,000 downloads. That counter measures file pulls and can include repeated downloads by one user, but it indicates substantial interest in a local model aimed at autonomous code repair and multi-turn tool use.
A 25-issue tie with Opus
SWE-bench-Live tests whether an agent can resolve real GitHub issues by inspecting a repository, editing code, and running tests. On the maintainers’ selected 25-problem slice, Tiel produced the following results:
| Model | Issues solved | Solve rate |
|---|---|---|
| Tiel-Coder-35B-A3B | 12 of 25 | 48% |
| Opus 4.6, medium | 12 of 25 | 48% |
| Nail | 9 of 25 | 36% |
| Ornith-1.5-35B-A3B | 8 of 25 | 32% |
| Sonnet 5, medium | 8 of 25 | 32% |
Each additional solve moves the score by four percentage points, so the Opus tie supports a narrow task comparison rather than broad model parity. The repository provides no confidence intervals for this slice. Its broader table places Tiel first among the listed MoE quantizations near its size, with dense 27B models and Opus 5 scoring higher.
Tiel completed an attempt in a median of 8.6 minutes, compared with Nail’s 7.2 minutes. Its mean was lower at 12.3 minutes versus 15.7, indicating fewer or shorter long-running attempts in this sample.
Chat scores rise to 67.2
Claw-Eval measures performance across multi-turn conversations, where a model must retain context and respond coherently over several exchanges. Across 114 scored conversations per model, Tiel outperformed both Nail and its Ornith base: