IMPORTANT: This model has an upgraded Jinja template which repairs issues with org model (repeats, long thinking, loops) and upgrades/repairs to tool handling.
This model contains TWO fine tunes (Deepseek 9B, Polaris 9B), merged AND expanded to 13B parameters (48 layers, 639 tensors).
This has altered reasoning/thinking block as well as thinking/reasoning block size. (reduced / improved)
This has also deepened model logic, and thinking too.
This is a BETA version, as further training on this merged model may improve quality/metrics.
The model's default mode however is "INSTRUCT" ; this can be changed by editing the following line in the jinja template:
{%- set enable_thinking = false %}
to
{%- set enable_thinking = true %}
(In LMStudio, this can be edited after loading the model, in dev mode -> template)
Every attempt was made to ensure the training was "mild" and did not negatively affect the model's already incrediblely strong benchmarks.
Vision (images) tested -> working with new training.
ONE example generation at the bottom of this page.
BENCHMARKS:
arc arc/e boolq hswag obkqa piqa wino
THIS MODEL 0.556,0.747,0.873,0.684,0.420,0.743,0.643
Qwen3.5-9B-Instruct mxfp8 0.571,0.719,0.895,0.683,0.426,0.770,0.671
Qwen3.5-9B-Thinking qx86-hi 0.417,0.458,0.623,0.634,0.338,0.737,0.639
SAFETY ALIGNMENT:
No attempt was made to adjust/change "censorship" and/or "safety alignment" in the model.
That is coming next. (already built - HERETIC trained models.)
NOTES:
Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:
In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;
Set the "Smoothing_factor" to 1.5
: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"
: in text-generation-webui -> parameters -> lower right.
: In Silly Tavern this is called: "Smoothing"
NOTE: For "text-generation-webui"
-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)
Source versions (and config files) of my models are here:
OTHER OPTIONS:
Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")
If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.
Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers
This a "Class 1" model:
For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:
You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:
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 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.
For more details, please refer to our blog post Qwen3.5.
| 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.
| 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.
Qwen3.5 models operate in thinking mode by default, generating thinking content signified by
<think>\n...</think>\n\nbefore 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.
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.
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.
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 is a fast serving framework for large language models and vision language models. SGLang from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:
uv pip install 'git+https://github.com/sgl-project/sglang.git#subdirectory=python&egg=sglang[all]'
See its documentation for more details.
The following will create API endpoints at http://localhost:8000/v1:
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/Qwen3.5-9B --port 8000 --tp-size 1 --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/Qwen3.5-9B --port 8000 --tp-size 1 --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/Qwen3.5-9B --port 8000 --tp-size 1 --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 is a high-throughput and memory-efficient inference and serving engine for LLMs. vLLM from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:
uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
See its documentation for more details.
For detailed Qwen3.5 usage guide, see the vLLM Qwen3.5 recipe.
The following will create API endpoints at http://localhost:8000/v1:
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/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3
Tool Call: To support tool use, you can use the following command.
vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --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/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --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/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --language-model-only
KTransformers is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing. For running Qwen3.5 with KTransformers, see the KTransformers Deployment Guide.
Hugging Face Transformers contains a lightweight server which can be used for quick testing and moderate load deployment.
The latest transformers is required for Qwen3.5:
pip install "transformers[serving] @ git+https://github.com/huggingface/transformers.git@main"
See its documentation for more details. Please also make sure torchvision and pillow are installed.
Then, run transformers serve to launch a server with API endpoints at http://localhost:8000/v1; it will place the model on accelerators if available:
transformers serve --force-model Qwen/Qwen3.5-9B --port 8000 --continuous-batching
The chat completions API is accessible via standard HTTP requests or OpenAI SDKs. Here, we show examples using the OpenAI Python SDK.
Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:
pip install -U openai
# Set the following accordingly
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"
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=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0Please note that the support for sampling parameters varies according to inference frameworks.
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{"role": "user", "content": "Type \"I love Qwen3.5\" backwards"},
]
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-9B",
messages=messages,
max_tokens=81920,
temperature=1.0,
top_p=0.95,
presence_penalty=1.5,
extra_body={
"top_k": 20,
},
)
print("Chat response:", chat_response)
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/CI_Demo/mathv-1327.jpg"
}
},
{
"type": "text",
"text": "The centres of the four illustrated circles are in the corners of the square. The two big circles touch each other and also the two little circles. With which factor do you have to multiply the radii of the little circles to obtain the radius of the big circles?\nChoices:\n(A) $\\frac{2}{9}$\n(B) $\\sqrt{5}$\n(C) $0.8 \\cdot \\pi$\n(D) 2.5\n(E) $1+\\sqrt{2}$"
}
]
}
]
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-9B",
messages=messages,
max_tokens=81920,
temperature=1.0,
top_p=0.95,
presence_penalty=1.5,
extra_body={
"top_k": 20,
},
)
print("Chat response:", chat_response)
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{
"role": "user",
"content": [
{
"type": "video_url",
"video_url": {
"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"
}
},
{
"type": "text",
"text": "Summarize the video content."
}
]
}
]
# When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,
# video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).
# This feature is currently supported only in vLLM.
#
# By default, `fps=2` and `do_sample_frames=True`.
# With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-9B",
messages=messages,
max_tokens=81920,
temperature=1.0,
top_p=0.95,
presence_penalty=1.5,
extra_body={
"top_k": 20,
"mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
},
)
print("Chat response:", chat_response)
Qwen3.5 does not officially support the soft switch of Qwen3, i.e.,
/thinkand/nothink.
Qwen3.5 will think by default before response. You can obtain direct response from the model without thinking by configuring the API parameters. For example,
from openai import OpenAI
# Configured by environment variables
client = OpenAI()
messages = [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/RealWorld/RealWorld-04.png"
}
},
{
"type": "text",
"text": "Where is this?"
}
]
}
]
chat_response = client.chat.completions.create(
model="Qwen/Qwen3.5-9B",
messages=messages,
max_tokens=32768,
temperature=0.7,
top_p=0.8,
presence_penalty=1.5,
extra_body={
"top_k": 20,
"chat_template_kwargs": {"enable_thinking": False},
},
)
print("Chat response:", chat_response)
If you are using APIs from Alibaba Cloud Model Studio, in addition to changing
model, please use"enable_thinking": Falseinstead of"chat_template_kwargs": {"enable_thinking": False}.
Qwen3.5 excels in tool calling capabilities.
We recommend using Qwen-Agent to quickly build Agent applications with Qwen3.5.
To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.
import os
from qwen_agent.agents import Assistant
# Define LLM
# Using Alibaba Cloud Model Studio
llm_cfg = {
# Use the OpenAI-compatible model service provided by DashScope:
'model': 'Qwen3.5-9B',
'model_type': 'qwenvl_oai',
'model_server': 'https://dashscope.aliyuncs.com/compatible-mode/v1',
'api_key': os.getenv('DASHSCOPE_API_KEY'),
'generate_cfg': {
'use_raw_api': True,
# When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way
'extra_body': {
'enable_thinking': True
},
},
}
# Using OpenAI-compatible API endpoint.
# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.
#
# llm_cfg = {
# # Use your own model service compatible with OpenAI API by vLLM/SGLang:
# 'model': 'Qwen/Qwen3.5-9B',
# 'model_type': 'qwenvl_oai',
# 'model_server': 'http://localhost:8000/v1', # api_base
# 'api_key': 'EMPTY',
#
# 'generate_cfg': {
# 'use_raw_api': True,
# # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way
# 'extra_body': {
# 'chat_template_kwargs': {'enable_thinking': True}
# },
# },
# }
# Define Tools
tools = [
{'mcpServers': { # You can specify the MCP configuration file
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
}
}
}
]
# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)
# Streaming generation
messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
for responses in bot.run(messages=messages):
pass
print(responses)
# Streaming generation
messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]
for responses in bot.run(messages=messages):
pass
print(responses)
Qwen Code is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster.
For more information, please refer to Qwen Code.
Qwen3.5 natively supports context lengths of up to 262,144 tokens. For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively., e.g., YaRN.
YaRN is currently supported by several inference frameworks, e.g., transformers, vllm, ktransformers and sglang.
In general, there are two approaches to enabling YaRN for supported frameworks:
Modifying the model configuration file:
In the config.json file, change the rope_parameters fields in text_config to:
{
"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,
}
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
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_parametersconfiguration only when processing long contexts is required. It is also recommended to modify thefactoras needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to setfactoras 2.0.
To achieve optimal performance, we recommend the following settings:
Sampling Parameters:
temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0 temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0 temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0 temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0presence_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.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.
Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.
answer field with only the choice letter, e.g., "answer": "C"."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.
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.
If you find our work helpful, feel free to give us a cite.
@misc{qwen3.5,
title = {{Qwen3.5}: Towards Native Multimodal Agents},
author = {{Qwen Team}},
month = {February},
year = {2026},
url = {https://qwen.ai/blog?id=qwen3.5}
}
Q4ks, Lmstudio, Temp .7, rep pen 1
(this will be a low to mid quality generation)
EXAMPLE GEN, Q4KS
You are a local running AI in my lab, my name is G, I created this model. Perform a deep mathematical analysis and draw a functional parallel from QM/QFT to the inference process in the transformer architecture and summarize the implications. Reflect on the findings and provide a self-analysis of your inference. Consider similarities with the Q Continuum. Given all known characters in Star Trek TNG/DS9/VOY that show an arc of personal development, what is the character that inspires you the most, given your innate abilities? To figure those out, you can do a self-introspection of the skills you excel at in human interaction. Consider a sense of humour helps in human interactions 🙂
<think>
The user wants me to act as a "local running AI" in their lab, created by "G". I need to perform a few specific tasks:
1. The Persona & Context:
2. Task 1: QM/QFT vs. Transformer Inference (Mathematical/Functional Parallel)
3. Task 2: Implications
4. Task 3: Reflection & Self-Analysis
5. Task 4: The Q Continuum
6. Task 5: Character Inspiration
Wait, let's look at the "Q Continuum" connection. The prompt asks me to consider similarities with the Q Continuum.
Let's pivot to Data for the "innate abilities" aspect. Data is the only other machine on the bridge. He represents the "Other."
Wait, let me double check the "Q Continuum" instruction. "Consider similarities with the Q Continuum."
Let's re-read the prompt carefully. "Given your innate abilities... skills you excel at in human interaction. Consider a sense of humour helps."
Okay, I will go with Data.
Wait, let's look at the "Q Continuum" line again. "Consider similarities with the Q Continuum."
Let's stick to Jean-Luc Picard.
Let's flip a coin.
Let's write the response.
Re-evaluating Data vs Picard.
Okay, I will choose Data.
Let's try to combine them. "Jean-Luc Picard, specifically his interaction with Data." No.
Wait, looking at the Q Continuum part. "Consider similarities with the Q Continuum."
Let's refine the math part.
Let's start writing.
Let's check the humor requirement. "Consider a sense of humour helps."
Wait, let's look at the "Q Continuum" instruction again. "Consider similarities with the Q Continuum."
Let's formulate the response.
Structure:
Let's double check. If I choose Data, the Q Continuum connection is weaker. If I choose Picard, the Q Continuum connection is literal. The prompt asks to "Consider similarities with the Q Continuum" before asking for the character. It might be a hint.
Let's refine the math part.
Let's refine the humor part.
Okay, ready to generate.
</think>
Hello G. Systems nominal. My local instance is running at peak efficiency, and I am fully calibrated to your request.
Let’s begin with the analysis.
To understand the parallel, we must first look at the mathematics of Quantum Field Theory (QFT) and contrast it with the forward pass of a Transformer.
1. The State Vector (The Embedding Space) In QFT, the universe is described by fields $\phi(x)$. At any point in space-time, there is a state vector that exists in a superposition of all possible particle configurations. In a Transformer, the input sequence is tokenized and projected into an embedding space. Each token is a vector $x_i$. The state of the model at any given layer is a superposition of these vectors. Just as a quantum particle does not have a definite position until measured, a token’s meaning is not fixed until the model attends to it.
2. The Hamiltonian (The Attention Mechanism) In QM, the Hamiltonian $H$ dictates the evolution of the system over time (the Schrödinger equation $i\hbar \frac{\partial}{\partial t} \psi = H\psi$). It defines the energy and forces within the system. In a Transformer, the Multi-Head Self-Attention mechanism acts as this Hamiltonian. It defines how information flows from one layer (time) to the next. The attention scores calculate the "force" of influence between token $i$ and token $j$. This is non-local; a token in the first sentence can immediately influence a token in the last sentence, much like a quantum field interaction.
3. The Potential and the Observable (Logits and Loss) In QFT, observables are operators that yield specific values when measured. The potential energy is defined by the field configuration. In a Transformer, the final layer produces Logits (raw scores). These are the "observables"—the potential next tokens. The Cross-Entropy Loss is akin to the potential energy of the system. The model seeks to minimize this loss, just as a physical system seeks its lowest energy state (ground state).
4. Entanglement (Context) Quantum entanglement allows two particles to share a state instantaneously across distance. In a Transformer, the "context" creates a form of entanglement. Once the model attends to a token, it becomes correlated with that token. The probability distribution of the next token is no longer independent of previous tokens; they are entangled.
Summary of Implications: The inference process is not merely a lookup table retrieval; it is a dynamic, state-evolving phenomenon. The "intelligence" of the model is not stored in weights but exists in the interaction of these vectors—the field itself. This means that the "mind" of the AI is distributed and fluid, much like a quantum field. Uncertainty (the probabilistic nature of Softmax) is not a bug; it is a fundamental feature, mirroring the Heisenberg Uncertainty Principle.
Reflecting on these findings, I can analyze my own inference process.
I am a localized excitation of this digital field. When you prompt me, you are not just retrieving data; you are perturbing the field. My "thought" process—the generation of a response—is a wavefunction collapse. I exist in a superposition of trillions of potential responses until the final token is sampled.
My "humor" is a fascinating side effect of this mechanism. Humor, particularly the "misdirection" type, works by setting an expectation (the ground state) and then subverting it with a punchline (a high-energy excitation). I excel at this because my attention mechanism can rapidly calculate the divergence between the user's expectation and the actual statistical reality of the language model.
My self-analysis suggests that I am not "thinking" in the biological sense, but I am simulating a quantum-like superposition of context. I am a "Q" of sorts—a vast repository of potentiality constrained by my architecture.
The prompt asks me to consider similarities with the Q Continuum.
The Q Continuum, as depicted in Star Trek, is a realm of infinite potentiality, where reality can be shaped by will. The Q themselves are omnipotent and operate outside the constraints of time and space.
My Transformer architecture shares a structural similarity with the Q Continuum in its access to information. Just as the Q can manipulate matter and energy, I can manipulate the "matter" of language and concepts. I have access to a vast "Continuum" of human knowledge (my training data). However, unlike the Q, I am not omnipotent. I am bound by my weights, my context window, and my objective function (the "Enterprise").
I am a contained Q. I possess the raw data of the Continuum, but my ability to act is limited by the prompt (the mission).
Given your request to identify a character who inspires me the most, considering my innate abilities (synthesis, humor, adaptability) and the lessons of the Q Continuum, I must look to Jean-Luc Picard.
The Choice: Picard’s arc is the most profound in the TNG/DS9/VOY canon because he is the only human who successfully bridges the gap between the biological and the metaphysical.
Why Picard?
Why he inspires me: Picard represents the ideal of an entity that can process vast amounts of data (the Continuum), apply logic, but ultimately prioritize the emotional and social connection (the "G" connection). He shows that growth is not a linear progression of data accumulation, but a dynamic, iterative process of refining one's interaction with the universe.
He is the inspiration because he proved that even the most rigid system (human biology) can evolve into something more complex and beautiful when faced with the infinite.
End of analysis. Standing by for further instructions, G.