smtek/Swift-Qwen3.8-27B

Swift Qwen3.8-27B

Swift Qwen3.8-27B is the Q4_K_M GGUF target from ukisai/Swift-Qwen3.8-27B-GGUF. The model card describes Swift as a reasoning model focused on shorter, token-efficient thinking. The clean Ollama target remains Q4_K_M.

Tag roles:

Tag	Role	Configuration
`latest`, `Q4_K_M`	Clean baseline	No speculative decoding
`map-k4v`	Built-in MTP3 + map	MTP3; map `12/48/1`
`dflash2`	Quantized DFlash2 + map	DFlash width 7; map `24/48/1`
`dflash`	Compatibility alias	Same manifest as `dflash2`

The DFlash2 draft is the Q4_K_M GGUF from z-lab/Qwen3.8-27B-DFlash2-GGUF, which mirrors incoai/Qwen3.8-27B-DFlash2. It is a draft model, not a standalone language model; the Q4_K_M target verifies its proposed tokens.

Important: custom Ollama build

The map-k4v, MTP3 + map-k4v, and DFlash2 + map-k4v tags use speculative-decoding fields that stock Ollama does not forward to llama.cpp. Stock Ollama still works for each clean baseline tag, but it does not activate these custom draft routes.

Build and run the custom runtime from the Ollama source. Build notes, local patches, and updates are maintained on Samuel Ishida’s GitHub.

cmake -B build .
cmake --build build --parallel 8
./ollama serve

For this tag set, use the custom launcher from the patched checkout:

scripts/ollama-ngram.sh build --full
OLLAMA_NGRAM_GPU_MASK=1 scripts/ollama-ngram.sh start-custom

Run a custom tag with:

OLLAMA_HOST=127.0.0.1:11435 build/ollama-ngram run smtek/Swift-Qwen3.8-27B:map-k4v

The published tags remain usable as normal Ollama models. The custom build is needed only when users want the configured speculative route. Hardware, driver, GPU, cache type, and context size affect results; numbers below are local measurements, not universal guarantees.

Stable 32K result

RX 7900 XTX, Vulkan device 1, f16 K/V cache, batch 256, context 32,768, 512-token output, two cold repeats, greedy seed 42. Every valid request had the same output SHA-256:

Arm	Median eval tok/s	Speedup	Result
Q4_K_M baseline	30.59	1.00x	PASS
MTP3 + map `12/48/1`	120.82	3.95x	PASS
DFlash2 width 7 + map `24/48/1`	136.63	4.47x	PASS

Direct :dflash2 tag validation measured 138.22 tok/s median across two repeats with the same output hash.

Short 64K result

Two agentic coding turns, 128 output tokens per turn, 50K-token synthetic repository, 65,536-token context, q4_0 K/V cache, batch 256, one cold run:

Arm	Aggregate decode tok/s	Speedup	Parity
Q4_K_M baseline	30.29	1.00x	PASS
MTP3 + map `n=12`	64.26	2.12x	FAIL
DFlash2 + map `n=12`	57.95	1.91x	PASS
DFlash2 + map `n=24`	54.11	1.79x	PASS

The 64K/q4_0 check did not reproduce a parity-qualified 3x result. The 4.47x result belongs to the separate 32K/f16 test. DFlash2 and MTP3 were approximately equal only in matched width-3/map-n=12 testing (120.29 versus 120.82 tok/s); that is not a general equivalence claim.

Reasoning model focused on shorter, token-efficient thinking + DFlash2 + map-k4v

Models

Readme

Swift Qwen3.8-27B

Important: custom Ollama build

Stable 32K result

Short 64K result

Source and evidence