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.

• 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

  bash

  1apt-get update 2apt-get install pciutils build-essential cmake curl libcurl4-openssl-dev -y 3git clone https://github.com/ggml-org/llama.cpp 4cmake llama.cpp -B llama.cpp/build \ 5 -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON 6cmake --build llama.cpp/build --config Release -j --clean-first --target llama-cli llama-mtmd-cli llama-server llama-gguf-split 7cp llama.cpp/build/bin/llama-* llama.cpp
  bash

  1export LLAMA_CACHE="unsloth/Qwen3.5-397B-A17B-MTP-GGUF" 2./llama.cpp/llama-server \ 3 -hf unsloth/Qwen3.5-397B-A17B-MTP-GGUF:UD-Q4_K_XL \ 4 -ngl 99 -c 8192 -fa on -np 1 \ 5 --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.

Mar 5 'Final' Update:

• All GGUFs now use our new imatrix data. See some improvements in chat, coding, long context, and tool-calling use-cases.
• GGUFs updated with an improved quantization algorithm.
• Updated with fixed chat template for improved tool-calling & coding performance!
• See our new benchmarks for 122B-A10B here.

Fine-tuning and RL Qwen3.5

• You can also fine-tune and perform reinforcement learning (RL) on all Qwen3.5 models with Unsloth via our free Colab notebooks.
• Read our Qwen3.5 fine-tuning guide for tips, VRAM requirements, code and more here.

Qwen3.5-397B-A17B

[!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.

Qwen3.5 Highlights

• 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.

Model Overview

• Type: Causal Language Model with Vision Encoder
• Training Stage: Pre-training & Post-training
• Language Model
  • Number of Parameters: 122B in total and 10B activated
  • Hidden Dimension: 3072
  • Token Embedding: 248320 (Padded)
  • Number of Layers: 48
  • Hidden Layout: 16 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE))
  • Gated DeltaNet:
    • Number of Linear Attention Heads: 64 for V and 16 for QK
    • Head Dimension: 128
  • Gated Attention:
    • Number of Attention Heads: 32 for Q and 2 for KV
    • Head Dimension: 256
    • Rotary Position Embedding Dimension: 64
  • Mixture Of Experts
    • Number of Experts: 256
    • Number of Activated Experts: 8 Routed + 1 Shared
    • Expert Intermediate Dimension: 1024
  • 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

Vision Language

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.

Serving Qwen3.5

[!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. If you encounter out-of-memory (OOM) errors, consider reducing the context window. However, because Qwen3.5 leverages extended context for complex tasks, we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.

SGLang

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.
  python -m sglang.launch_server --model-path Qwen/Qwen3.5-397B-A17B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3
• Tool Use: To support tool use, you can use the following command.
  python -m sglang.launch_server --model-path Qwen/Qwen3.5-397B-A17B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder
• Multi-Token Prediction (MTP): The following command is recommended for MTP:
  python -m sglang.launch_server --model-path Qwen/Qwen3.5-397B-A17B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

vLLM

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.
  vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3
• Tool Call: To support tool use, you can use the following command.
  vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder
• Multi-Token Prediction (MTP): The following command is recommended for MTP:
  vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
• Text-Only: The following command skips the vision encoder and multimodal profiling to free up memory for additional KV cache:
  vllm serve Qwen/Qwen3.5-397B-A17B --port 8000 --tensor-parallel-size 8 --max-model-len 262144 --reasoning-parser qwen3 --language-model-only

KTransformers

Hugging Face Transformers

Using Qwen3.5 via the Chat Completions API

[!Tip] We recommend using the following set of sampling parameters for generation

• Thinking mode for general tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Thinking mode for precise coding tasks (e.g. WebDev): temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
• Instruct (or non-thinking) mode for general tasks: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Instruct (or non-thinking) mode for reasoning tasks: temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0

Please note that the support for sampling parameters varies according to inference frameworks.

Text-Only Input

Image Input

Video Input

Instruct (or Non-Thinking) Mode

[!Important] Qwen3.5 does not officially support the soft switch of Qwen3, i.e., /think and /nothink.

[!Note] If you are using APIs from Alibaba Cloud Model Studio, in addition to changing model, please use "enable_thinking": False instead of "chat_template_kwargs": {"enable_thinking": False}.

Agentic Usage

Qwen-Agent

Qwen Code

Processing Ultra-Long Texts

• Modifying the model configuration file: In the config.json file, change the rope_parameters fields in text_config to:
  json

  1{ 2 "mrope_interleaved": true, 3 "mrope_section": [ 4 11, 5 11, 6 10 7 ], 8 "rope_type": "yarn", 9 "rope_theta": 10000000, 10 "partial_rotary_factor": 0.25, 11 "factor": 4.0, 12 "original_max_position_embeddings": 262144, 13}
• Passing command line arguments:
  For vllm, you can use
  VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1010000
  For sglang and ktransformers, you can use
  SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1010000

[!NOTE] All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts. We advise modifying the rope_parameters configuration only when processing long contexts is required. It is also recommended to modify the factor as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set factor as 2.0.

Best Practices

1. Sampling Parameters:
   • We suggest using the following sets of sampling parameters depending on the mode and task type:
     • Thinking mode for general tasks:
       temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
     • Thinking mode for precise coding tasks (e.g., WebDev):
       temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
     • Instruct (or non-thinking) mode for general tasks:
       temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
     • Instruct (or non-thinking) mode for reasoning tasks:
       temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
   • For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.
2. Adequate Output Length: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 81,920 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.
3. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.
   • Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
   • Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the answer field with only the choice letter, e.g., "answer": "C"."
4. No Thinking Content in History: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.
5. Long Video Understanding: To optimize inference efficiency for plain text and images, the size parameter in the released video_preprocessor_config.json is conservatively configured. It is recommended to set the longest_edge parameter in the video_preprocessor_config file to 469,762,048 (corresponding to 224k video tokens) to enable higher frame-rate sampling for hour-scale videos and thereby achieve superior performance. For example,
   {"longest_edge": 469762048, "shortest_edge": 4096}
   Alternatively, override the default values via engine startup parameters. For implementation details, refer to: vLLM / SGLang.

Citation