Tiel-Inspired-Coder-9B-GGUF

Tiel-Inspired-Coder-9B (GGUF)

Tiel-Coder's methodology, applied to Ornith-1.5-9B. No fine-tuning — just an imatrix, a chat template, and dynamic quantization.

What this is

This is Ornith-1.5-9B re-quantized with a custom imatrix and carrying the Sharp chat template inside the GGUF — the same recipe used by Tiel-Coder-35B-A3B, scaled down to the 9B.

This is not a Tiel-Coder release. It is an independent reproduction of peculiar-ragdoll's methodology (imatrix + Sharp template, no weight changes) on the smaller Ornith-1.5-9B base. The name "Tiel-Inspired-Coder" makes that clear. Full credit to peculiar-ragdoll for the approach and the Sharp template. Full credit to ornith-ai for the base model.

The imatrix

The calibration corpus is ~49M characters drawn from eaddario/imatrix-calibration (MIT): about three quarters code_medium and one quarter combined_all_large, interleaved — the same mix Tiel-Coder 35B used.

The matrix was measured on a Q8_0 of the original BF16 weights, over 3,000 chunks of 512 tokens (~1.5M tokens). Final PPL estimate: 6.1217.

The imatrix file (Ornith-1.5-9B.imatrix.gguf) ships in this repo for anyone who wants to cut additional tiers.

The Sharp template

The Sharp chat template (qwen3.8-froggeric-v22.4.1) is embedded in every GGUF. It enables reasoning by default, supports configurable reasoning effort, and adds a terseness directive that keeps answers lean.

Reasoning-discipline template (chat_template.jinja)

A drop-in alternative of the embedded template, with an extra reasoning-discipline block injected into the system prompt whenever reasoning is enabled. It keeps the exact same structure, tooling, and effort handling — it only adds four rules that force the model to conclude:

• Reason ONLY as short bullet points; stop when the bullets contain the decision and plan.
• Never re-verify once the decision is made.
• If the spec contradicts itself, name the discrepancy in one bullet and commit to the standard interpretation.

Use it at runtime with llama.cpp by pointing at the file:

./build/bin/llama-server -m Tiel-Inspired-Coder-9B-UD-Q4_K_M.gguf \
    --jinja --chat-template-file chat_template.jinja

This avoids the degenerate "overthinking" repetition loops without changing the weights.

Available quants

All tiers were quantized from the BF16 source using the imatrix. The Sharp template was injected into each file's metadata.

How to run

hf download VitreousCut/Tiel-Inspired-Coder-9B-GGUF Tiel-Inspired-Coder-9B-UD-Q4_K_M.gguf --local-dir .
llama-server -m Tiel-Inspired-Coder-9B-UD-Q4_K_M.gguf -ngl 99 --jinja -c 32768

Sampling: temperature 0.2, top_p 0.95, top_k 20 for coding tasks.

Important: This model is a massive overthinker

The Sharp template enables reasoning by default. On simple problems the model may produce a few hundred tokens of reasoning before answering — fine. On complex or ambiguous problems, it can spiral into 32,000+ tokens of reasoning and never produce an answer at all, even with a 32k token budget.

This is the single biggest weakness of this build. In HumanEval testing, 3 out of 164 problems failed purely because the model spent its entire token budget thinking without emitting code — and no amount of additional budget fixed them. These are "incurable overthinkers" where the model gets stuck in a reasoning loop.

Recommended fix: use reasoning_effort: low

The Sharp template supports configurable reasoning effort. For most use cases, pass reasoning_effort: low via chat template kwargs:

llama-server -m Tiel-Inspired-Coder-9B-UD-Q4_K_M.gguf -ngl 99 --jinja -c 32768 \
  --reasoning-effort low

Or via API:

{
  "messages": [...],
  "chat_template_kwargs": {"reasoning_effort": "low"}
}

This keeps reasoning concise and focused. Reserve medium (default) or high for genuinely hard problems where you can afford a longer think.

HumanEval results

All runs on Q4_K_M, 3 parallel workers, temperature 0.2.

Primary run (reasoning_effort=medium, 16k token budget)

Error re-run (32k token budget, 131k context)

Re-ran the 6 error problems with 32k token budget and 131k context:

Adjusted Pass@1 with re-run: 95.1% (156/164)

The 3 remaining overthinking failures are problems where the model loops endlessly in its reasoning block. No amount of token budget fixes them — this is a behavioral issue that would require fine-tuning to address.

Comparison: reasoning_effort=low

low is 2x faster and half the tokens, but loses ~4 points of accuracy. The sweet spot for quality is medium with a big budget; the sweet spot for speed is low.

Baseline comparison

Ornith-1.5-9B base typically scores in the 72-78% range on HumanEval at FP16. The imatrix + Sharp template + reasoning budget pushed it to 93.3% (95.1% adjusted) at 4-bit quantization.

Hardware

Tested on:

• CPU: Intel i7-6700K
• GPU: NVIDIA RTX 3090 (24GB VRAM)
• RAM: 32GB

All tiers run fully offloaded on a 3090 with room for 32k+ context.

Credits

• ornith-ai — the Ornith-1.5-9B weights (MIT)
• peculiar-ragdoll — the Tiel-Coder methodology, the Sharp chat template, and the imatrix recipe
• eaddario — the calibration corpora (MIT)
• froggeric — the template lineage Sharp builds on
• llama.cpp — llama-quantize / llama-imatrix / llama-server

MIT, inheriting Ornith-1.5's license.