Why this matters
Local development workflows increasingly need models that fix real codebases, read screenshots/stack traces, and sustain multi-turn debugging without cloud latency or data exfiltration. Tiel is positioned exactly for that trade: it re-quantizes Ornith-1.5-35B with a tailored importance matrix and a concise chat template to favor actionable answers and coding edits over long-form exam-style reasoning.
Key Capabilities
- Focused coding performance: on a 25-problem SWE-bench-Live sample Tiel fixes 12 problems, matching Opus 4.6 (medium) and outperforming its base and several sibling builds. That reflects practical bug-fix effectiveness rather than synthetic exam scores.
- Conversation-first tuning: scores higher on multi-turn conversation benchmarks (Claw-Eval ~67.2) by providing more direct answers and fewer clarifying questions, which speeds agentic workflows where the user prefers decisive fixes.
- Size/quantization tiers for local GPUs: tiers range from ~12.3 GB (Q2) up to ~38.5 GB (Q8); the recommended benchmarked tier is a 4-bit ~22.4 GB file that fits comfortably on 32 GB cards and aims for a balance of accuracy and context headroom.
- Vision-enabled debugging: inherits Ornith-1.5's vision tower (mmproj-BF16), letting you point the model at screenshots, stack traces or failing tests; the projector ships at BF16 and is unchanged by text quantization.
- Quantization and template trade-offs: uses a shipped imatrix calibrated mostly on coding data and the Sharp chat template to shorten answers; this improves coding throughput but reduces exam-style MMLU Pro scores (73.7 at 4-bit) versus some peers.
Who it's for and trade-offs
Great fit if you want a locally runnable model that: performs practical code fixes on real repos, handles multimodal debugging (screenshots/logs), and prioritizes concise, action-oriented responses for agentic coding workflows. It is tuned to be token- and runtime-efficient for single-GPU setups.
Look elsewhere if you need top exam/knowledge benchmark scores or maximal open-domain trivia accuracy: builds focused on MMLU-style reasoning (e.g., Nail) outperform Tiel on exam benchmarks. Also pick the MTP variant if you rely on multi-token speculative decoding; this repo's tiers omit the multi-token head to save ~0.9 GB unless you explicitly choose the -MTP release.
Where it fits
Tiel sits between research-grade dense models and larger cloud APIs: it trades some general knowledge for faster, steadier attempt times on debugging tasks and stronger multi-turn coding dialogue. Use it when local, responsive, code-focused assistance and multimodal inspection matter more than peak test-set exam scores.
