npario/Dirk-Qwen3.8-27B-GGUF

Dirk is the Qwen3.8-27B that gets straight to the point.

With our Sharp chat template, MTP, and vision baked in, the model answers lean and stays on-task out of the box. No template wrangling: download, point llama.cpp at it, go. If you want it to think deeper, set the effort level through chat_template_kwargs:

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

Levels: low, medium, xhigh — high is accepted but is an alias for xhigh, not a step below it. Omit it for Dirk's lean default (medium). Turn thinking off entirely with "enable_thinking": false.

What it is

• Base: Qwen/Qwen3.8-27B, a dense 27B vision-language model (vision preserved).
• Quant: Unsloth's Dynamic 3.0 UD GGUF quants — their current generation, not an older ladder — which keep the model's MTP (nextn) head: runtimes with multi-token-prediction speculative decoding can use it for faster generation.
• Template: the Sharp chat template (Qwen 3.8-aware) — froggeric's fixed Qwen template plus a terseness system prompt, and turning off the xhigh thinking default. IQ4_XS carries v22.3.2, which adds an opt-out (chat_template_kwargs: {"terse": false}); the older tiers carry v22.3.1 and render identically unless you use that opt-out. It is byte-swapped into the GGUF metadata; the weights and the MTP tensors are untouched.

The only thing Dirk changes versus the stock quant is the template. Same weights, asked better.

Proven on Nail and Dagger

Dirk is new, but the template is not. The identical terseness edit, measured on Dagger's base (ThinkingCap-27B, same weights, only the template swapped):

	stock template	Sharp template	change
Claw-Eval, answer component	59.3	66.7	+7.4
Claw-Eval answer tokens	5393	2217	−59%
MMLU-Pro tokens per correct answer	1601	1248	−22%

Roughly: the same answers in a bit over half the words, with accuracy moving up. That is what Dirk inherits — and its own SWE-bench-Live and MMLU-Pro numbers, shown above, bear it out.

Thinking effort

Stock Qwen3.8-27B forces reasoning_effort=xhigh on every call — always-on maximum-effort reasoning. Dirk removes that default, so it runs at the model's native medium effort: in both the official and Unsloth templates, medium is the setting that injects no reasoning instruction (only xhigh and low add one), and Dirk simply leaves it there. So Dirk thinks at the baseline and answers terse, instead of being pushed to the ceiling on every request. Set reasoning_effort yourself (low, medium, xhigh; high maps to xhigh), per request, through chat_template_kwargs — the OpenAI-style top-level reasoning_effort field is dropped by llama.cpp and oMLX, so it must go there (see the JSON example above).

Run it

file	size	notes
Dirk-Qwen3.8-27B-UD-Q2_K_XL.gguf	9.8 GB	the 12 GB card pick — smallest tier; 2-bit gives up real accuracy, so prefer a bigger one wherever it fits
Dirk-Qwen3.8-27B-UD-IQ3_XXS.gguf	10.9 GB	fits 16 GB with room to spare, and a 12 GB card at shorter context — better than Q2_K_XL wherever it fits
Dirk-Qwen3.8-27B-UD-Q3_K_XL.gguf	13.1 GB	3-bit with headroom to spare on 16 GB; prefer IQ4_XS below unless you need the extra ~1 GB for context
Dirk-Qwen3.8-27B-UD-IQ4_XS.gguf	14.3 GB	the 16 GB pick — 4-bit quality with room for real context, where Q4_K_S leaves almost none
Dirk-Qwen3.8-27B-UD-Q4_K_S.gguf	15.4 GB	tight 4-bit; useful when Q4_K_XL will not fit alongside your context
Dirk-Qwen3.8-27B-UD-Q4_K_XL.gguf	17.6 GB	start here — the 24 GB-card default; best size/quality balance
Dirk-Qwen3.8-27B-UD-Q5_K_XL.gguf	20.9 GB	the recommended 24 GB pick — dynamic + imatrix-calibrated, and small enough to leave real room for context
Dirk-Qwen3.8-27B-UD-Q6_K.gguf	22.0 GB	6-bit — the largest that still fits 24 GB, with tighter headroom than UD-Q5_K_XL
Dirk-Qwen3.8-27B-UD-Q6_K_XL.gguf	25.3 GB	near-max quality; wants ~32 GB
Dirk-Qwen3.8-27B-UD-Q8_K_XL.gguf	31.5 GB	8-bit, effectively lossless

Every file carries the Sharp template and the MTP (nextn) head, and all share mmproj-F16.gguf for vision — you need only one copy of it. Every tier is an Unsloth Dynamic 3.0 (UD) quant.

Let llama.cpp fetch it — pass a :quant tag from the table (:Q4_K_XL, :IQ4_XS, :Q6_K_XL, …). The tag is required: this repo has no Q4_K_M, so a bare -hf with no tag falls back to the wrong file. The mmproj rides along in the manifest, so vision works from the same tag — no second download.

# text — auto-downloads to llama.cpp's own cache (24 GB-card default shown)
llama-server   -hf peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF:Q4_K_XL -ngl 99   # or llama-cli
# vision — same tag; the mmproj is pulled automatically
llama-mtmd-cli -hf peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF:Q4_K_XL -ngl 99 --image photo.jpg

Prefer to keep the files yourself? Download explicitly, then point -m at the local path:

hf download peculiar-ragdoll/Dirk-Qwen3.8-27B-GGUF Dirk-Qwen3.8-27B-UD-Q4_K_XL.gguf \
  mmproj-F16.gguf --local-dir Dirk
llama-cli      -m Dirk/Dirk-Qwen3.8-27B-UD-Q4_K_XL.gguf -ngl 99                              # text
llama-mtmd-cli -m Dirk/Dirk-Qwen3.8-27B-UD-Q4_K_XL.gguf --mmproj Dirk/mmproj-F16.gguf -ngl 99  # vision

llama.cpp applies the embedded Sharp template automatically — nothing to pass.

Driving it from a coding agent? Add --reasoning-format deepseek to llama-server. It returns the model's <think> block in the OpenAI reasoning_content field instead of inline in content, so the agent never sees raw thinking tokens in the text stream. Current llama.cpp already defaults to this (--reasoning-format auto is defined as "same as deepseek"), so it is a no-op on a recent build and insurance on an older one. Just don't pass --reasoning-format none — that is the one that leaves the tags inline.

Pick your weapon

Qwen3.8-27B may be the new intelligence density frontier for local models that run on consumer hardware, but the already battle-tested Dagger and Nail, joined by the newer TielCoder, each have their own use cases, in an arsenal that contains all four.

• Nail-35B-A3B generates tokens 3–4× faster than 27B models, while still being very good at routine coding, debugging, knowledge work, and many other kinds of tasks — which means that for tasks that aren't too hard for it, it writes the unit test and regression test, and implements the feature in the time it takes 3.8-27B to get out of the gate. Reach for Nail when you need volume routine work done right and fast.
• TielCoder-35B-A3B is the dedicated coder: Nail's 35B-A3B speed class, rebuilt on Ornith-1.5 with the Sharp template and pointed at one job. It fixes 12 of 25 on SWE-bench-Live — level with Opus 4.6, four clear of Sonnet 5 (medium) — at the lowest mean time per attempt of the 35B-A3B family. It pays for that in general knowledge: 73.7 on MMLU-Pro against Nail's 84.0. Reach for TielCoder when the work is code; reach for Nail when the same session also has to know things.
• Dagger-27B is — unlike 3.8-27B — specifically tuned to minimize the number of thinking tokens while sacrificing minimal accuracy, which might still give it the advantage in speed-to-answer and multi-turn stamina under the context ceiling. Reach for Dagger when you need a session to survive 100 turns.
• Dirk-27B is what you reach for when the task is genuinely hard and you want the strongest local answer without filler — accepting that Nail reaches an answer faster on work it can handle, and that a marathon session running 100 turns under the context ceiling is Dagger's domain, not Dirk's.

Dagger, Nail and TielCoder might still be your go-to workhorses for long and short tasks within their ability bands, due to their advantage in speed and stamina.

Credits

• Qwen — the Qwen3.8-27B weights.
• Unsloth — the UD-Q4_K_XL Dynamic quant (MTP-preserving) this repo redistributes.
• froggeric — the fixed chat template the Sharp template builds on.

Apache-2.0, matching upstream.