Qwen3.5-DeltaCoder-9B-GGUF
v1.1-DPO — Now with DPO alignment for improved code correctness and self-verification. If you downloaded before March 28, 2026, please re-pull to get v1.1-DPO.
GGUF quantizations of Qwen3.5-DeltaCoder-9B for use with llama.cpp, Ollama, LM Studio, and other GGUF-compatible inference engines.
What's New in v1.1-DPO
- DPO alignment on 4,519 preference pairs from AceCode-V2-122K
- Self-correcting behavior — model now detects and fixes its own bugs rather than submitting incorrect code
- Improved code correctness — trained to prefer passing solutions over failing ones
- Same tool-call reliability as v1 — SFT improvements preserved through two-stage merge
Available Quantizations
| File | Quant | Size | Notes |
|---|---|---|---|
DeltaCoder-9B-v1.1-DPO-Q2_K.gguf | Q2_K | ~3.6 GB | Smallest, lowest quality |
DeltaCoder-9B-v1.1-DPO-Q3_K_S.gguf | Q3_K_S | ~4.0 GB | |
DeltaCoder-9B-v1.1-DPO-Q3_K_M.gguf | Q3_K_M | ~4.4 GB | |
DeltaCoder-9B-v1.1-DPO-Q3_K_L.gguf | Q3_K_L | ~4.6 GB | |
DeltaCoder-9B-v1.1-DPO-Q4_0.gguf | Q4_0 | ~3.2 GB | |
DeltaCoder-9B-v1.1-DPO-Q4_K_S.gguf | Q4_K_S | ~5.0 GB | |
DeltaCoder-9B-v1.1-DPO-Q4_K_M.gguf | Q4_K_M | ~5.5 GB | Recommended |
DeltaCoder-9B-v1.1-DPO-Q5_K_S.gguf | Q5_K_S | ~6.1 GB | |
DeltaCoder-9B-v1.1-DPO-Q5_0.gguf | Q5_0 | ~6.1 GB | |
DeltaCoder-9B-v1.1-DPO-Q5_K_M.gguf | Q5_K_M | ~6.5 GB | |
DeltaCoder-9B-v1.1-DPO-Q6_K.gguf | Q6_K | ~7.3 GB | |
DeltaCoder-9B-v1.1-DPO-Q8_0.gguf | Q8_0 | ~9.4 GB | Near-lossless |
DeltaCoder-9B-v1.1-DPO-BF16.gguf | BF16 | ~17.9 GB | Full precision |
Recommended Quant
- Low VRAM (8GB): Q4_K_M
- Mid VRAM (12GB): Q5_K_M or Q6_K
- High VRAM (16GB+): Q8_0
- Full precision: BF16
Training Lineage
Qwen/Qwen3.5-9B-Base
└─ Qwen/Qwen3.5-9B (instruction tuned)
└─ Jackrong/Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2
(SFT on Claude 4.6 Opus reasoning traces)
└─ danielcherubini/Qwen3.5-DeltaCoder-9B (v1 SFT — tool-call reliability)
(LoRA SFT on CoderForge-Preview)
└─ danielcherubini/Qwen3.5-DeltaCoder-9B v1.1-DPO ← this model
(DPO on AceCode-V2-122K preference pairs)
Recommended Sampling Settings
| Parameter | Value |
|---|---|
| temperature | 0.6 |
| top_k | 20 |
| top_p | 0.95 |
| min_p | 0.0 |
| presence_penalty | 0.0 |
| repeat_penalty | 1.0 |
[!WARNING] Do not use temperature below 0.5 — low temperatures cause deterministic looping in multi-turn agentic use.
KV Cache Quantization
| Context Length | KV Cache | VRAM (Q4_K_M) | Generation Speed |
|---|---|---|---|
| 102,400 | f16/q4_0 | ~8.5 GB | ~111 tok/s |
| 131,072 | f16/q4_0 | ~9.1 GB | ~110 tok/s |
# llama.cpp / ik_llama.cpp flags
-ctk f16 -ctv q4_0
Usage
Ollama
ollama create deltacoder -f Modelfile
Example Modelfile:
FROM ./DeltaCoder-9B-v1.1-DPO-Q5_K_M.gguf
llama.cpp
./llama-server -m DeltaCoder-9B-v1.1-DPO-Q5_K_M.gguf -ngl 999 -c 131072 -ctk f16 -ctv q4_0 -fa 1 --jinja
LM Studio
Download any GGUF file and load it directly in LM Studio.
Benchmarks
| Model | HumanEval | HumanEval+ | Terminal-Bench Easy |
|---|---|---|---|
| Jackrong Qwen3.5-9B-v2 (base) | 53.7% | — | — |
| DeltaCoder-9B v1 (temp=0.6) | 50.6% | 49.4% | 2/4 (50%) |
| DeltaCoder-9B v1.1-DPO (temp=0.6) | TBD | TBD | 2/4 (50%)* |
*v1.1-DPO timed out on 2 tasks that v1 answered incorrectly — behavioral improvement confirmed, running with extended timeout.
Acknowledgements
- Unsloth for Qwen3.5 training support
- Together AI for the CoderForge dataset
- TIGER Lab for AceCode-V2-122K
- Jackrong for the reasoning distillation
- Qwen for the base model