Qwen3.5-9B-MTP-GGUF

Read our How to Run Qwen3.5 MTP Guide!#

To run Qwen3.5 locally - Read our Guide!#

Unsloth Dynamic 2.0 achieves superior accuracy & outperforms other leading quants.

https://github.com/unslothai/unsloth/ https://discord.gg/unsloth https://unsloth.ai/docs/models/qwen3.5

NEW: MTP speculative decoding for ~1.5-2x faster generation - build llama.cpp from the MTP PR branch - has been merged 16th May 2026!

Note: llama.cpp renamed --spec-type mtp to --spec-type draft-mtp on 2026-05-13

apt-get update
apt-get install pciutils build-essential cmake curl libcurl4-openssl-dev -y
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build \
    -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON
cmake --build llama.cpp/build --config Release -j --clean-first --target llama-cli llama-mtmd-cli llama-server llama-gguf-split
cp llama.cpp/build/bin/llama-* llama.cpp

export LLAMA_CACHE="unsloth/Qwen3.5-9B-MTP-GGUF"
./llama.cpp/llama-server \
    -hf unsloth/Qwen3.5-9B-MTP-GGUF:UD-Q4_K_XL \
    -ngl 99 -c 8192 -fa on -np 1 \
    --spec-type draft-mtp --spec-draft-n-max 6

Set -DGGML_CUDA=OFF for CPU/Metal. -np > 1 and --mmproj are not yet supported with MTP.

Developer Role Support so Qwen3.5 can work in Codex, OpenCode and more!

Tool calling improvements: Makes parsing nested objects to make tool calling succeed more.

• You can now also fine-tune the model locally with Unsloth.
• Read our Qwen3.5 fine-tuning guide here.

Qwen3.5-9B#

Image: image

Qwen Chat

Note. This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency.

Qwen3.5 Highlights#

Qwen3.5 features the following enhancement:

• Unified Vision-Language Foundation: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.
• Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.
• Scalable RL Generalization: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability.
• Global Linguistic Coverage: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding.
• Next-Generation Training Infrastructure: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.

Image: Benchmark Results

For more details, please refer to our blog post Qwen3.5.

Model Overview#

• Type: Causal Language Model with Vision Encoder
• Training Stage: Pre-training & Post-training
• Language Model
  • Number of Parameters: 9B
  • Hidden Dimension: 4096
  • Token Embedding: 248320 (Padded)
  • Number of Layers: 32
  • Hidden Layout: 8 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
  • Gated DeltaNet:
    • Number of Linear Attention Heads: 32 for V and 16 for QK
    • Head Dimension: 128
  • Gated Attention:
    • Number of Attention Heads: 16 for Q and 4 for KV
    • Head Dimension: 256
    • Rotary Position Embedding Dimension: 64
  • Feed Forward Network:
    • Intermediate Dimension: 12288
  • LM Output: 248320 (Padded)
  • MTP: trained with multi-steps
• Context Length: 262,144 natively and extensible up to 1,010,000 tokens.

Benchmark Results#

Language#

| | GPT-OSS-120B | GPT-OSS-20B | Qwen3-Next-80B-A3B-Thinking | Qwen3-30BA3B-Thinking-2507 | Qwen3.5-9B | Qwen3.5-4B

| Knowledge & STEM

| MMLU-Pro | 80.8 | 74.8 | 82.7 | 80.9 | 82.5 | 79.1

| MMLU-Redux | 91.0 | 87.8 | 92.5 | 91.4 | 91.1 | 88.8

| C-Eval | 76.2 | 71.4 | 89.7 | 87.4 | 88.2 | 85.1

| SuperGPQA | 54.6 | 48.5 | 60.8 | 56.8 | 58.2 | 52.9

| GPQA Diamond | 80.1 | 71.5 | 77.2 | 73.4 | 81.7 | 76.2

| Instruction Following

| IFEval | 88.9 | 88.2 | 88.9 | 88.9 | 91.5 | 89.8

| IFBench | 69.0 | 65.1 | 61.5 | 51.5 | 64.5 | 59.2

| MultiChallenge | 45.3 | 40.1 | 51.3 | 46.5 | 54.5 | 49.0

| Long Context

| AA-LCR | 50.7 | 30.7 | 51.7 | 49.0 | 63.0 | 57.0

| LongBench v2 | 48.2 | 45.6 | 48.0 | 44.8 | 55.2 | 50.0

| Reasoning & Coding

| HMMT Feb 25 | 90.0 | 76.7 | 73.7 | 63.1 | 83.2 | 74.0

| HMMT Nov 25 | 90.0 | 81.8 | 81.2 | 73.8 | 82.9 | 76.8

| LiveCodeBench v6 | 82.7 | 74.6 | 68.7 | 66.0 | 65.6 | 55.8

| OJBench | 41.5 | 36.3 | 29.7 | 25.1 | 29.2 | 24.1

| General Agent

| BFCL-V4 | -- | -- | 49.7 | 42.4 | 66.1 | 50.3

| TAU2-Bench | -- | -- | 57.4 | 41.9 | 79.1 | 79.9

| VITA-Bench | -- | -- | 29.5 | 14.1 | 29.8 | 22.0

| DeepPlanning | -- | -- | 0.4 | 4.9 | 18.0 | 17.6

| Multilingualism

| MMMLU | 78.2 | 69.7 | 81.3 | 78.4 | 81.2 | 76.1

| MMLU-ProX | 74.5 | 67.3 | 73.6 | 69.1 | 76.3 | 71.5

| NOVA-63 | 51.1 | 48.7 | 53.3 | 52.5 | 55.9 | 54.3

| INCLUDE | 74.0 | 65.3 | 78.3 | 74.4 | 75.6 | 71.0

| Global PIQA | 84.1 | 79.8 | 83.5 | 80.2 | 83.2 | 78.9

| PolyMATH | 54.0 | 30.9 | 62.4 | 52.6 | 57.3 | 51.1

| WMT24++ | 74.4 | 67.8 | 57.4 | 69.3 | 72.6 | 66.6

| MAXIFE | 83.7 | 80.1 | 79.9 | 77.4 | 83.4 | 78.0

• TAU2-Bench: we follow the official setup except for the airline domain, where all models are evaluated by applying the fixes proposed in the Claude Opus 4.5 system card.

• MMLU-ProX: we report the averaged accuracy on 29 languages.
• WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL.
• MAXIFE: we report the accuracy on English + multilingual original prompts (totally 23 settings).
• Empty cells (--) indicate scores not yet available or not applicable.

Vision Language#

| | GPT-5-Nano-2025-08-07 | Gemini-2.5-Flash-Lite | Qwen3-VL-30B-A3B | Qwen3.5-9B | Qwen3.5-4B

| STEM and Puzzle

| MMMU | 75.8 | 73.4 | 76.0 | 78.4 | 77.6

| MMMU-Pro | 57.2 | 59.7 | 63.0 | 70.1 | 66.3

| MathVision | 62.2 | 52.1 | 65.7 | 78.9 | 74.6

| Mathvista(mini) | 71.5 | 72.8 | 81.9 | 85.7 | 85.1

| We-Math | 62.5 | 32.1 | 70.0 | 75.2 | 75.4

| DynaMath | 78.0 | 69.9 | 80.1 | 83.6 | 83.3

| ZEROBench | 1.0 | 1.0 | 0.0 | 3.0 | 3.0

| ZEROBench_sub | 22.2 | 19.2 | 23.7 | 31.1 | 26.3

| VlmsAreBlind | 66.7 | 68.4 | 72.5 | 93.7 | 92.6

| BabyVision | 14.4 | 17.5 | 18.6 | 28.6/25.8 | 16.0/19.1

| General VQA

| RealWorldQA | 71.8 | 72.2 | 77.4 | 80.3 | 79.5

| MMStar | 68.6 | 69.1 | 75.5 | 79.7 | 78.3

| MMBenchEN-DEV-v1.1 | 80.3 | 82.7 | 88.9 | 90.1 | 89.4

| SimpleVQA | 46.0 | 54.1 | 54.3 | 51.2 | 43.4

| HallusionBench | 58.4 | 64.5 | 66.0 | 69.3 | 65.0

| Text Recognition and Document Understanding

| OmniDocBench1.5 | 55.9 | 79.4 | 86.8 | 87.7 | 86.2

| CharXiv(RQ) | 50.1 | 56.1 | 56.6 | 73.0 | 70.8

| MMLongBench-Doc | 31.8 | 46.5 | 47.4 | 57.7 | 54.2

| CC-OCR | 58.9 | 72.9 | 77.8 | 79.3 | 76.7

| AI2D_TEST | 81.9 | 85.7 | 86.9 | 90.2 | 89.6

| OCRBench | 75.3 | 82.5 | 83.9 | 89.2 | 85.0

| Spatial Intelligence

| ERQA | 45.8 | 44.3 | 45.3 | 55.5 | 54.0

| CountBench | 80.0 | 79.2 | 90.0 | 97.2 | 96.3

| RefCOCO(avg) | -- | -- | 89.3 | 89.7 | 88.1

| EmbSpatialBench | 74.2 | 66.1 | 80.6 | 83.0 | 81.3

| RefSpatialBench | 12.6 | 11.2 | 54.2 | 58.5 | 54.6

| LingoQA | 57.0 | 17.8 | 62.0 | 80.4 | 74.4

| Hypersim | -- | -- | 11.4 | 13.5 | 12.5

| Nuscene | -- | -- | 10.3 | 11.8 | 9.9

| Video Understanding

| VideoMME(w sub.) | 71.7 | 74.6 | 79.9 | 84.5 | 83.5

| VideoMME(w/o sub.) | 66.2 | 72.7 | 73.3 | 78.4 | 76.9

| VideoMMMU | 63.0 | 69.2 | 75.0 | 78.9 | 74.1

| MLVU | 69.2 | 78.5 | 78.9 | 84.4 | 82.8

| MVBench | -- | -- | 72.0 | 74.4 | 71.2

| LVBench | -- | 60.9 | 59.2 | 70.0 | 66.4

| MMVU | 63.1 | 65.3 | 66.1 | 67.8 | 64.9

| Visual Agent

| ScreenSpot Pro | -- | -- | 60.5 | 65.2 | 60.3

| OSWorld-Verified | -- | -- | 30.6 | 41.8 | 35.6

| AndroidWorld | -- | -- | 55.0 | 57.8 | 58.6

| Tool Calling

| TIR-Bench | 18.5 | 21.5 | 22.5 | 45.6/31.9 | 38.9/29.9

| V* | 68.1 | 69.6 | 83.2 | 90.1/88.5 | 84.3/86.4

| Medical VQA

| SLAKE | 57.0 | 65.0 | 68.8 | 79.0 | 76.1

| PMC-VQA | 37.8 | 48.8 | 51.5 | 57.9 | 55.5

| MedXpertQA-MM | 26.7 | 35.3 | 35.5 | 49.9 | 42.9

• MathVision: our model's score is evaluated using a fixed prompt, e.g., "Please reason step by step, and put your final answer within \boxed{}." For other models, we report the higher score between runs with and without the \boxed{} formatting.
• BabyVision: scores reported as "with CI / without CI".
• TIR-Bench and V*: scores reported as "with CI / without CI".
• Empty cells (--) indicate scores not yet available or not applicable.

Quickstart#

Important. Qwen3.5 models operate in thinking mode by default, generating thinking content signified by <think>\n...</think>\n\n before producing the final responses. To disable thinking content and obtain direct response, refer to the examples here.

For streamlined integration, we recommend using Qwen3.5 via APIs. Below is a guide to use Qwen3.5 via OpenAI-compatible API.

Serving Qwen3.5#

Qwen3.5 can be served via APIs with popular inference frameworks. In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.5 models.

Important. Inference efficiency and throughput vary significantly across frameworks. We recommend using the latest framework versions to ensure optimal performance and compatibility. For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.

Important. The model has a default context length of 262,144 tokens.

The publisher's README is longer; it is cut here. Read the rest at the source.