Qwen3.8-3.6-27B-blend GGUF

Unofficial JetBrains derivative. This model was prepared by JetBrains using Qwen models developed by Alibaba Cloud. This release is not affiliated with, sponsored by, or endorsed by Alibaba Cloud or Alibaba Group.

Tasks completed versus output tokens on JetBrains’ 100-task internal coding benchmark. Reasoning settings differ between configurations; see the article for methodology and context.

Read Making Local AI Smarter and Faster for the Junie Local release story, coding evaluations, token-efficiency results, and MTP runtime experiments.

Available formats: BF16 · GGUF · MLX 4-bit · MTP MLX 4-bit.

IQ3_S, Q3_K_XL, IQ4_XS, Q4_K_M, and Q5_K_M conversions of JetBrains/Qwen3.8-3.6-27B-blend, pinned to revision f1a19acf58aa8caf7a0a507c9083245f79944dcc.

The base model is a 50/50 combination of Qwen3.6-27B and Qwen3.8-27B, created by linearly interpolating their checkpoint parameters. The configuration, tokenizer, processor, and chat template are retained from Qwen3.8-27B. Exact source revisions and merge settings are recorded in merge-manifest.json. See conversion-manifest.json for conversion settings and runtime versions.

The MTP head is embedded in each main GGUF. A separate drafter file is not required. Either vision GGUF can be used with any of the five main-model quantizations; choose one with --mmproj for image input.

Run

With the prebuilt llama binary (llama.app and other distributors):

llama serve -hf JetBrains/Qwen3.8-3.6-27B-blend-GGUF

Add the native MTP head for faster decoding:

llama serve -hf JetBrains/Qwen3.8-3.6-27B-blend-GGUF --spec-type draft-mtp --spec-draft-n-max 3

Or the DFlash2 drafter (see the sidecar section below):

llama serve -hf JetBrains/Qwen3.8-3.6-27B-blend-GGUF --spec-type draft-dflash --spec-draft-n-max 7

-hf reuses the cached download on repeat runs. To pin a quantization, append the tag: -hf JetBrains/Qwen3.8-3.6-27B-blend-GGUF:Q4_K_M. For vision input, pass --mmproj with one of the two projector files. API requests should retain the source sampling settings: temperature 1.0, top-p 0.95, top-k 20, with thinking enabled.

The equivalent long-form llama-server invocation, for explicit control over every knob:

llama-server \
  --model Qwen3.8-3.6-27B-blend-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.8-3.6-27B-blend-BF16.gguf \
  --alias Qwen3.8-3.6-27B-blend \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --n-gpu-layers 99 --ctx-size 8192 --parallel 2 \
  --jinja --reasoning on --reasoning-format deepseek \
  --chat-template-kwargs '{"enable_thinking":true,"preserve_thinking":true}' \
  --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0 --repeat-penalty 1 \
  --host 127.0.0.1 --port 8080

The MTP flow above was tested on llama.cpp revision 64e9bceb2c3a856efed96feda784a50947049feb. DFlash2 support and sidecar discovery are upstream llama.cpp (PRs #30548, #30663, #30927) and are included in current prebuilt binaries; the original IQ3_S/Q4_K_M/Q5_K_M quants were validated on the tested revision, the DFlash2 sidecar smoke on a current build.

Conversion

The original IQ3_S, Q4_K_M, and Q5_K_M quants were produced directly from one BF16 GGUF intermediate, using llama.cpp's default quantization recipes, without an importance matrix. They were not converted from previously quantized MLX weights.

python convert_hf_to_gguf.py /path/to/pinned-bf16-source \
  --outtype bf16 --outfile model-BF16.gguf
./build/bin/llama-quantize model-BF16.gguf model-IQ3_S.gguf IQ3_S 10
./build/bin/llama-quantize model-BF16.gguf model-Q4_K_M.gguf Q4_K_M 10
./build/bin/llama-quantize model-BF16.gguf model-Q5_K_M.gguf Q5_K_M 10
python convert_hf_to_gguf.py /path/to/pinned-bf16-source \
  --mmproj --outtype bf16 --outfile mmproj-BF16.gguf

Environment: Python 3.12.14, PyTorch 2.11.0, Transformers 5.14.0, NumPy 1.26.4; llama.cpp built with Metal on an Apple M5 Max with 128 GiB RAM. Full conversion and validation settings are in conversion-manifest.json.

Validation

All five quants retain all 866 main-file tensors, including all 15 MTP tensors, with one next-token-prediction layer declared in the GGUF metadata.

The original IQ3_S, Q4_K_M, and Q5_K_M passed local text, image, streaming, two-request concurrency, invalid-request handling, client-disconnect cleanup, and post-cancellation recovery checks on Metal. The image test correctly identified a red square and blue circle. Text responses included separate reasoning content with thinking enabled.

On the first arithmetic smoke request, Q4_K_M accepted 98/114 drafted tokens (86.0%) and Q5_K_M accepted 78/93 (83.9%). These are individual smoke results, not comparative quality or speed benchmarks. No BF16 parity or broad quality evaluation is claimed.

IQ3_S was added on 2026-09-15 using the same BF16 intermediate and llama.cpp revision. It passed the same functional checks with thinking and native MTP enabled. Its arithmetic smoke request accepted 113/144 drafted tokens (78.5%). This is a smoke-test observation, not a quality or performance benchmark. The exact artifact SHA256, input SHA256, and validation results are recorded in conversion-manifest.json.

Calibrated IQ4_XS — 2026-09-30

Qwen3.8-3.6-27B-blend-IQ4_XS.gguf is a 14.25 GB mixed-precision quantization, approximately 15.2% smaller than the published Q4_K_M. Its tensor-type assignments follow Unsloth's Qwen3.8 UD-IQ4_XS artifact, applied to the blend's BF16 weights with a new importance matrix computed on 32 calibration windows of 8,192 tokens (262,144 tokens of text). The matrix is the 32-chunk checkpoint of a longer calibration run. llama.cpp wrote it before accumulating the final 512-token micro-batch, so 495 of its 496 entries cover 261,632 tokens and one covers 262,144. This is a JetBrains quantization using the published tensor assignments, not an official Unsloth release or a reproduction of its recipe-selection process. The filename denotes the recipe; individual tensors use several quantization formats.

All 866 tensor shapes and types were verified. Text-only BF16-relative KL validation was performed with MTP disabled. No additional fine-tuning was performed. The functional checks reported above for the original quants were not run at release; on 2026-10-17 the published IQ4_XS file passed the same text, image, streaming, two-request concurrency, invalid-request, client-disconnect, and recovery checks on Metal with thinking and native MTP enabled. Its arithmetic smoke request accepted 87/102 drafted tokens (85.3%). This is a smoke-test observation, not a quality or performance benchmark.

The upload copy replaces local calibration path labels in GGUF metadata with portable names; its tensor payload is byte-identical to the evaluated candidate. See quantization-IQ4_XS.json for checksums and calibration settings, and IQ4_XS.tensor-types.txt for the exact tensor assignments.

Calibrated Q3_K_XL — 2026-10-17

Qwen3.8-3.6-27B-blend-Q3_K_XL.gguf is a 13.15 GB mixed-precision quantization, approximately 21.8% smaller than the published Q4_K_M. Its tensor-type assignments follow Unsloth's Qwen3.8 UD-Q3_K_XL artifact, applied to the blend's BF16 weights with the same importance matrix used for IQ4_XS (32 calibration windows of 8,192 tokens; 495 of 496 entries cover 261,632 tokens, as described above). This is a JetBrains quantization using the published tensor assignments, not an official Unsloth release or a reproduction of its recipe-selection process. The filename denotes the recipe; individual tensors use several quantization formats, and the GGUF file-type field reads Q3_K_L, the base type passed to llama-quantize.

All 866 tensor names, shapes, and types match the Unsloth reference. Text-only BF16-relative KL validation was performed with MTP disabled. The file passed the same text, image, streaming, two-request concurrency, invalid-request, client-disconnect, and recovery checks on Metal as the other quants, with thinking and native MTP enabled; its arithmetic smoke request accepted 211/255 drafted tokens (82.7%). This is a smoke-test observation, not a quality or performance benchmark. No additional fine-tuning was performed.

The upload copy replaces local calibration path labels in GGUF metadata with portable names; its tensor payload is byte-identical to the evaluated build. See quantization-Q3_K_XL.json for checksums and calibration settings, and Q3_K_XL.tensor-types.txt for the exact tensor assignments.

Q8_0 vision projector — 2026-10-05

mmproj-Qwen3.8-3.6-27B-blend-Q8_0.gguf is 629,247,584 bytes (0.63 GB), 32.4% smaller than the 931,146,464-byte BF16 projector. It was quantized directly from the existing BF16 projector with the same pinned llama.cpp revision:

llama-quantize mmproj-Qwen3.8-3.6-27B-blend-BF16.gguf \
  mmproj-Qwen3.8-3.6-27B-blend-Q8_0.gguf Q8_0 10

All 334 tensor names and shapes are preserved: 83 tensors use Q8_0, 27 feed-forward down-projection matrices use F16, and 224 F32 tensors remain byte-identical. The F16 fallback is llama.cpp's default because these matrices have width 4,304, which is not divisible by Q8_0's block size of 32. No calibration dataset or importance matrix was used.

Both BF16 and Q8_0 projectors passed the same synthetic image smoke with the published Q4_K_M main model on Metal, with thinking and native MTP enabled: each correctly identified a red square followed by a blue circle. This is a functional smoke test, not a broad vision-quality comparison. See quantization-mmproj-Q8_0.json for checksums, tensor assignments, and validation settings. The original BF16 projector remains available.

DFlash2 drafter sidecars — 2026-10-26

Added two DFlash2 drafter sidecars for llama.cpp's --spec-type draft-dflash, named with the dflash- prefix llama.cpp uses to discover sidecars next to the main model:

Both are conversions of the DFlash2 draft model for this blend (v0.3, training step 500; weights SHA-256 fbee4cc4790035d0a4d4d7cece25f41a0f8d629b4c630d19be86a994ee90a477), fine-tuned from the public Qwen 3.8 DFlash2 drafter on Junie coding trajectories. The BF16 drafter GGUF was produced with llama.cpp's convert_hf_to_gguf.py --target-model-dir (81 drafter tensors; embed_tokens and lm_head are pulled from the blend target at load time, so the drafter file deliberately contains none). The Q4_K_M sidecar was quantized from that BF16 GGUF with the JetBrains llama.cpp fork (commit 8e778da57); it is a fresh quantization, not the file measured in the 2026-10-07 decode bench (different llama.cpp build, SHA-256 8466a3f5fe9c276426924b8941753c848e2750f3e84f37acd2b07e92d7a71b20).

llama.cpp discovers the sidecar next to the main model, so no separate draft download is needed:

llama serve -hf JetBrains/Qwen3.8-3.6-27B-blend-GGUF --spec-type draft-dflash --spec-draft-n-max 7

Smoke on Metal (M5 Max, 128 GB): the Q4_K_M sidecar loaded (block_size 8, mask_token_id 248070, n_extract 5), and a 512-token thinking-on request completed at 44.2 tok/s over 159 verification steps with per-position accepted tokens 53/19/11/9/3/0/0. Functional smoke only, not a quality or speed benchmark. For reference, a Q4_K_M build of the same drafter (different llama.cpp build) served the 2026-10-07 decode bench at 24.5 tok/s vs 13.6 baseline without drafting (1.8×) at n_max 7 on the Windows ARM bench host. The standalone draft repository remains the recommended artifact for vLLM and dflash[local], whose loaders require a root config.json.

License and attribution

Distributed under the Apache License, Version 2.0. The original Qwen LICENSE, including Copyright 2026 Alibaba Cloud, is retained unchanged.

Copyright © 2026 JetBrains s.r.o. This notice applies only to original material contributed by JetBrains. Upstream Qwen material remains subject to its original copyright and attribution notices.

See NOTICE for attribution and CHANGES.md for a description of the modifications and the affected files.

Note on acceptable use

This model is distributed under the Apache License, Version 2.0. This notice does not modify the License. Users are reminded that using this model, or its outputs, to infringe the copyright or other intellectual property rights of third parties may violate applicable law, independent of the terms of this License.

Training and upstream information

No additional training or fine-tuning was performed; IQ4_XS and Q3_K_XL use calibration data only for post-training quantization, and its preparation consisted of the checkpoint merge and, where applicable, the extraction, format conversion, and quantization described above and in CHANGES.md.

For upstream information, see the official Qwen3.6-27B model card and Qwen3.8-27B model card. Qwen also publishes a Training Data Summary, which describes training data for models powering qwen.ai generally; it does not identify the exact training datasets for these two checkpoints.