Tiel Repacks Ornith-1.5 to Match Claude Opus on Local Coding
A community re-quantization of Ornith-1.5-35B-A3B revives its multi-token prediction head, matching Opus 4.6 on SWE-bench-Live at 23 GB.
Takeaways
- Tiel-Coder re-quantizes Ornith-1.5-35B-A3B with a working multi-token-prediction draft head.
- Fixes 12 of 25 SWE-bench-Live problems at 4-bit, matching Opus 4.6 medium.
- Speculative decoding via
--spec-type draft-mtpyields up to 1.22x throughput on llama.cpp. - Scores 67.2 on Claw-Eval multi-turn, beating Nail (60.5) and the Ornith base (65.3).
- Trade-off: MMLU-Pro drops to 73.7 versus Nail's 84.0 due to terser Sharp template.
- MIT-licensed GGUFs from 12.7 GB (Q2) to 38.8 GB (Q8), plus shipped imatrix.
Tiel model card packages Ornith-1.5 as llama.cpp-compatible GGUF files with a repaired multi-token-prediction head, a new importance matrix, and a terser chat template called Sharp. The community release requires no additional fine-tuning and is available under the MIT license inherited from Ornith.
The model card reports 12 fixes across a 25-problem SWE-bench-Live sample, matching Opus 4.6 at its medium setting. Tiel ran at 4-bit precision from a 22.7 GB file, placing it within reach of local systems with 24 or 32 GB of combined GPU and CPU memory capacity.
| Model | Problems fixed | Fix rate |
|---|---|---|
| Tiel | 12 | 48% |
| Opus 4.6, medium | 12 | 48% |
| Nail | 9 | 36% |
| Ornith-1.5 | 8 | 32% |
| Sonnet 5, medium | 8 | 32% |
A 25-task sample can expose meaningful regressions or gains, but it cannot establish a broad ranking among models. The reported comparison also lacks an ablation that isolates the effects of the template, quantization, and other evaluation settings.
How the draft head was repaired
Multi-token prediction, or MTP, adds a small head that drafts future tokens while the main model verifies them. Accepted drafts let llama.cpp process several tokens during a forward pass, reducing the number of sequential decoding steps.
Ornith’s initial quantization ladder contained an untrained MTP block. Every matrix had a standard deviation of 0.020 and kurtosis of 3.00, a distribution consistent with fresh random initialization. Such weights produce unusable drafts, so the head was removed from that release.
According to the repository history, Ornith replaced the block on 2026-08-23 by uploading a shard with learned MTP weights. The replacement showed kurtosis of 25.1, a 98-sigma outlier, and kurtosis of 825 in the