JackOD-9B-Coder — GGUF (DFlash)

ManniX-ITA/JackOD-9B-Coder v3 quantized for DFlash speculative decoding: every quant tier of JackOD-9B-Coder-MTP-GGUF with the built-in MTP (NextN) layer removed, plus a separate DFlash drafter that does the drafting instead.

Use this repo if your runtime drafts with DFlash. Use JackOD-9B-Coder-MTP-GGUF if it drafts with the model's own MTP head, or if you do not use speculative decoding at all. The two repos hold the same quantized weights; only the speculative-decoding path differs.

JackOD-9B-Coder is a merge over Qwen/Qwen3.5-9B built for multi-turn agentic coding — DeltaCoder-9B 0.55 / Ornith-1.5-9B 0.30 / Coder-Flash 0.15, omnimerge_v2. The recipe, the evaluations and the full description are on the model card.

What is in this repo

• 27 quant tiers, JackOD-9B-Coder-MTP-DFlash-<TIER>.gguf. Each is the v3 -MTP-GGUF file of the same tier with the NextN block stripped (omnimergekit/scripts/gguf_strip_nextn.py): 442 → 427 tensors, block_count 33 → 32, nextn_predict_layers removed, the per-layer arrays truncated to 32. Every kept tensor was verified byte-identical to the -MTP-GGUF file it came from (--verify), so the quantized weights — and the v3 imatrix they were calibrated with, sha256 feb286c42be30760e92fbb218dfc1ceb4be07a1cb01634993072d1f2ecc9dd80 — are exactly those of the MTP repo. Every file has a .sha256 sidecar.
• The drafter, Qwen3.5-9B-DFlash.Q8_0.gguf (1,383,768,224 bytes), redistributed unmodified from AtomicChat/Qwen3.5-9B-DFlash-GGUF at c7c69923d77a9aa0004b410d3a7a3dbb64591d41 (apache-2.0; base z-lab/Qwen3.5-9B-DFlash), sha256 27b9d18e605aea9c50ef506e4b63921e0dc624cc91c63c60617a8295f153d436 — the provenance line is in Qwen3.5-9B-DFlash.Q8_0.gguf.verified. It is a drafter for the Qwen3.5-9B family; JackOD shares that base's tokenizer and vocabulary.

Runtime support

Sampler

temperature 0.4 · top_p 0.95 · top_k 20 · min_p 0 · repeat_penalty 1.002 · repeat_last_n 112 · presence_penalty 1.05

The seed is not fixed. Worth trying: seed 42. In the author's testing it behaves better — no loops, and stellar on some tasks — but it is not reliable: it still fails on others. Treat it as something to test, not as part of the recommended setup.

Running it (llama.cpp)

llama-server -m JackOD-9B-Coder-MTP-DFlash-Q4_K_M.gguf \
  --spec-type draft-dflash -md Qwen3.5-9B-DFlash.Q8_0.gguf --spec-draft-n-max 4 \
  --jinja -c 32768 -ngl 99 \
  --temp 0.4 --top-p 0.95 --top-k 20 --min-p 0 \
  --repeat-penalty 1.002 --repeat-last-n 112 --presence-penalty 1.05

--jinja is mandatory for tool calling. Without it the tool grammar never comes from the model's chat template and every tool scenario degrades to prose.

Draft acceptance

Measured on every tier with llama.cpp 311d4211bf, --spec-type draft-dflash --spec-draft-n-max 4, on a 192-token and (for three tiers) a 768-token generation. These are greedy (temperature 0) measurements, taken before the sampler above was settled: acceptance under temperature 0.4 will be lower, and these numbers are a comparison between tiers, not a speedup promise. Draft acceptance depends on the content being generated.

Q3_K_S is the outlier (75.3% on the short run); every other tier accepts 84–95% of the drafted tokens on the short run.

Quants

The tiers, their sizes relative to the MTP files (each is 0.15–0.20 GB smaller: the NextN block is gone), the KLD-vs-F16 measurements and the ContribDynamic (CD-) bit-width maps are described on the -MTP-GGUF card. There is no F16 here; use the MTP repo's F16.

Verifying a download

sha256sum -c JackOD-9B-Coder-MTP-DFlash-Q4_K_M.gguf.sha256
sha256sum Qwen3.5-9B-DFlash.Q8_0.gguf   # 27b9d18e605aea9c50ef506e4b63921e0dc624cc91c63c60617a8295f153d436