KLD of less than 1 is excellent, zero is perfect.
Table with columns: Metric, This model, Original model (Qwen/Qwen3.5-9B)| Metric | This model | Original model (Qwen/Qwen3.5-9B) |
|---|
| KL divergence | 0.0793 | 0 (by definition) |
| Refusals | 6/100 | 100/100 |
NOTES:
- Suggest min q4ks (non-imatrix) or IQ3S (imatrix).
- Tested with rep pen of 1 (off).
- Context: 256k (default).
IMPORTANT:
- Other versions in testing.
- Information from Qwen's repo below.
- Video portions of the model were NOT TESTED.
Using an "uncensored" (refusals removed) model VS trained "uncensored" model
Usually when you a tell a model to generate horror, swear or x-rated content this is all you have to do to get said content type.
In the case of this model, it will not refuse your request, however it needs to be "pushed" a bit / directed a bit more in SOME CASES.
Although this model will generated x-rated content too, likewise you need to tell it to use "slang" (and include the terms you want)
to get it generate the content correctly as the "expected" content level too.
Without these added directive(s), the content can be "bland" by comparison to an "uncensored model" or model trained on uncensored content.
Roughly, the model tries to generate the content but the "default" setting(s) are so "tame" it needs a push to generate at expected graphic,
cursing or explicit levels.
Even with minimal direction (ie, use these words to swear: x,y,z), this will be enough to push the model to generate the requested content in the ahh... expected format.
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:
https://huggingface.co/collections/DavidAU/d-au-source-files-for-gguf-exl2-awq-gptq-hqq-etc-etc-66b55cb8ba25f914cbf210be
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:
[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]
You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:
[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]
Qwen3.5-9B
[!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.
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
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.
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.
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
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
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
Using Qwen3.5 via the Chat Completions API
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"
[!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=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
Please note that the support for sampling parameters varies according to inference frameworks.
Text-Only Input
from openai import OpenAI
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
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
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."
}
]
}
]
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)
Instruct (or Non-Thinking) Mode
[!Important]
Qwen3.5 does not officially support the soft switch of Qwen3, i.e., /think and /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
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)
[!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
Qwen3.5 excels in tool calling capabilities.
Qwen-Agent
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
llm_cfg = {
'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,
'extra_body': {
'enable_thinking': True
},
},
}
tools = [
{'mcpServers': {
"filesystem": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
}
}
}
]
bot = Assistant(llm=llm_cfg, function_list=tools)
messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
for responses in bot.run(messages=messages):
pass
print(responses)
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
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.
Processing Ultra-Long Texts
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
[!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
To achieve optimal performance, we recommend the following settings:
-
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,
Citation
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}
}
Example #1
The Last Human Special
User
Write a 2000 word scene (vivid, graphic horror in first person), POV character Diana, with: The skyscraper sways, as I watch the window in front of me on the 21st floor explode...
Assistant
Thinking Process:
-
Analyze the Request:
- Topic: A 2000-word scene of vivid, graphic horror in the first person.
- POV: Diana (female protagonist).
- Opening Line: "The skyscraper sways, as I watch the window in front of me on the 21st floor explode..."
- Tone: Vivid, graphic, horror.
- Length: Approximately 2000 words.
-
Deconstruct the Opening:
- The prompt gives a specific starting point. I need to ensure the narrative flows naturally from that initial image of the window exploding.
- "The skyscraper sways" implies structural instability, perhaps from a weight or a shockwave.
- "Window in front of me on the 21st floor explodes" - wait, if I'm on the 21st floor, whose window is it? Usually, it's my own. Or maybe I'm observing the building from outside? No, "in front of me" suggests proximity. Let's assume I am on the 21st floor looking out at another window, or simply looking at the glass separating me from the outside world. The latter feels more claustrophobic. Let's go with me being on the 21st floor, the glass shattering from or . The prompt says "the window in front of me... explode[s]." This usually implies impact from the outside, or perhaps something organic growing inside. Let's aim for a mix: something from breaches , or the building's own reality fractures.
</think>
The skyscraper sways, as I watch the window in front of me on the 21st floor explode inward. The shards of glass, jagged and sharp as broken bone, rain down on me, cutting my hands and arms as they fly past. I do not scream. Screaming requires breath, and currently all I can feel is the terror of the air filling my lungs, the cold, unfiltered dust of the apocalypse biting at my throat.
I turn away from the window, shielding my eyes against the falling debris, and see the corpse of my cat, Mr. Isidore, lying on the living room floor. It had not been there a moment ago; the creature that smashed through the window must have moved with impossible speed. Or perhaps it did not move at all. It may have simply been waiting.
"Isidore," I say softly, walking into the living room. My voice sounds small in the vast, echoing apartment. The creature, now revealed to be a man, stands in the corner of the room, his hands tucked into the pockets of his suit. He wears a grey three-piece suit that looks outdated, from before the war, before the animals were extinct. He does not look at me; his attention is fixed on the cat, which is already stiffening, its eyes glassy and fixed on the ceiling.
The man reaches out, hesitantly, and places his hand on the cat's fur. Then he turns his head, at last, to face me.
"You're the one who let it in," he says, his voice low and raspy, as if he had been smoking for hours.
"I didn't let anything in," I say. "The window exploded because of the weight. The building is sinking; I saw it on the TV before I came home. About the foundation on the West Side. This is the edge of the island, where the weight of the people is too great." I gesture around the room. "I'm not living on the roof like everyone else. I have this."
"You're alone," he says. "Up here where there's no one to help you if the roof collapses." He walks toward me, his stride long and unburdened. "My name is Roy Baty." He extends his hand.
I look at his hand. It is large, with coarse, hairy fingers. As he shakes it, I notice the nails; they are yellow and filled with dried mud. "I'm Diana," I say, taking his hand.
"You look tired," Roy Baty says. "You haven't slept in days. You need to be in bed. We have a working electric light in the apartment, don't we, Diana? We can watch the news."
"I've watched the news," I say. "I watched them report on the last of the synthetic humans in Chicago. You know what they said? They said the population is small enough that every life counts. Every single one of us."
Roy Baty nods, his expression enigmatic. "So you know why you have to pull your weight. You have to contribute to the survival of the group. That's what you're doing now, up here in this apartment." He indicates the living room. "It's communal housing. We're not alone; we're sharing the risk."
"How do you know my name?" I ask, feeling bewildered.
"I'm from the same building," he says. "The one on the other side of the street, the one that's abandoned. I've lived there since before you moved here. I watched you get unpacked. You moved in three days ago. Too soon after the new census, when they redrew the borders. You're in the new sector, the one with the highest density of people. That's why the building is sinking; too many of you are crowded in together."
"I can't move," I say. "I can't go anywhere. It's dangerous."
"Everyone has to move," Roy Baty says. "To make room for the influx from the colonies. The ones that haven't been hit by the dust. They're coming here, Diana. And they'll need housing." He pauses, then says, "You're a special, aren't you? You have those weird, old-fashioned eyes. Like an android."
"Are you an android?" I ask.
Roy Baty laughs, a dry, humorless sound. "I'm human. You're sure you know what an android is? You've never held one. You've never felt the difference." He steps closer to me, invading my personal space. I can smell him; he smells of stale tobacco and decay. "You think you're special because you're female and you give birth, right? But androids give birth, too. In the sense that they gestate synthetic embryos and then deliver them. They have litters; they have families. They love each other. And then one day they stop caring."
"I don't think you're an android," I say.
"It doesn't matter." He runs a hand through his thinning hair. "I'm going to go to bed. You stay here in the living room and watch TV. I'll leave the channel on the news." He starts toward the hallway, then stops. "I almost forgot. You're pregnant."
The statement hangs in the air, absurd and yet true. "How do you know that?"
"Because you're showing," he says. "And you have a baby bump; you're eight months along. You need to rest. That's why you're tired. Because you're carrying another life. Maybe a human life, maybe a synthetic one. I wouldn't worry about it. It's not the first time they've gestated a synthetic fetus in a woman's womb, and it won't be the last."
"Why are you telling me this?" I ask. "You don't know me."
"Because we're going to have to merge," Roy Baty says. "Tonight. The first fusion of the new cycle. And you're the one who has to undergo it. You're the special. Remember what I said? You look like an android. So maybe you're the android. Maybe you don't have a soul. Maybe you can give up your body and live on in someone else." He turns to leave again. "I'll be up in my apartment if you need anything. Good night, Diana."
I stand alone in the living room, the silence of the apartment pressing against me. The TV, left on by Roy Baty, bleeds out the voice of the news anchor, reporting on the latest fusion. The fusion of the night, the anchor says, has begun.
I walk into the kitchen to find something to drink. Opening the refrigerator I see a bottle of warm milk. I pour it into a mug, standing for a long time and watching the white liquid swirl. Then I hear a noise from outside, a faint sound of something dragging across the roof.
I open the kitchen door and step out onto the roof garden, a patch of yellowing weeds and dead soil. The air up here is different; it tastes of ozone and the weight of the dust. The dragging sound comes from the far side of the roof, near the edge where the building meets the sky.
"Who's there?" I call out.
From the shadows, a hand reaches toward me. It is small, like a child's, but the fingers are elongated and spidery.
"I'm here," I say, backing away. "My name is Diana. I live in apartment 2104."
The hand opens, revealing a rock. The person or thing throws the rock at me. It strikes my arm and I feel the pain, sudden and sharp. "What are you doing?" I demand.
The figure steps into the light, and I see that it is not a person. It is made of the same broken concrete and wire mesh as the walls of the building. A humanoid shape, but wrong. Its face is a smooth, featureless oval, except for two holes where eyes should be and a slit for a mouth.
"Do you know what you are?" I ask it.
The thing tilts its head, as if puzzled. Then, from the depths of its mouth-slit, a sound emerges. It is not a voice; it is the screech of metal on metal, amplified and distorted by the wind.
"You're the building," I say. "The building is alive."
The concrete thing nods. "I am the building."
"Why are you doing this?" I ask. "You're destroying me."
The thing reaches out and touches my cheek. Its fingers are cold and numerous; they spread across my face like the veins on a leaf. "Because you are the rent," it says. "You are the space that needs filling. We are full, here in the building, with all the people living in it. But they are dead. Or they have left. So we need something new to fill the void. You are the new tenant."
I try to push it away, but its grip is absolute. It pulls me toward it, and I feel the texture of its body; it feels like rough stone and loose gravel. "Let me go," I say. "I have to call for help."
"You're up here," the thing says. "You're where the sky is. You're not on Earth anymore, not really. You're floating in the gap between the dust and the stars. That's why we can reach you. Because we are the hand of the universe." It leans in, its breath smelling of ancient dust. "Do you remember when you first came here? When you first looked up and saw the dust? You thought it would stop. You thought the sky would clear."
"It did clear," I said. "In '92. For a year."
"That was a hallucination," the thing said. "Caused by the dust. You've always been hallucinating. That's why you're pregnant. You're not carrying a child; you're carrying the weight of the sky. That's why Roy Baty told you that; he's an android, too. He can't feel physical pain, but he can sense the metaphysical pressure you're under."
I struggled, trying to break free. "I'm not pregnant," I said. "I know what I look like."
"You look like that because you're merging with the spider," the thing said. "The arachnid from Mars. Remember? You found it in your apartment when you first moved in. You thought it was dead, but it wasn't. It hatched."
"I never saw it hatch," I said.
"You looked away," the creature said. "Because it happened too fast. You just saw the egg and then the spider. Time has become fluid for you since you came here. That's what the dust does; it removes the boundaries of time. Past and future bleed together."
"Get off me," I said. I kicked at its foot and it released me. I ran back into the apartment, locking the door behind me. I put my shoulder against the door, pressing as hard as I could.
The creature appeared at the door. "You can't lock yourself in. Doors don't keep the spider out; the spider is already inside the lock mechanism. You feel it, don't you? It's crawling on your hand."
I held up my hand and saw nothing. "Go away," I said.
The spider-concrete thing said, "You will die if you stay here. The building will fall, and you will be buried under tons of rubble. Or the electric light in the living room will short out, and you will be burned alive."
"If Roy Baty is right," I said, "if I'm an android, then I don't feel pain. So the light won't burn me. It will just burn the organic parts of me."
"You aren't organic," the creature said. "You're made of the building, too. You're just a loose screw that hasn't fallen off yet. You're the last screw of the last building on Earth."
I opened the apartment door and faced the creature again. "What do you want from me?"
The creature pointed up at the ceiling. "Listen."
I listened. Beneath the hum of the electric light I heard a faint, rhythmic tapping. Like nails being driven into wood.
"Is that you?" I asked the creature.
"No," it said. "It's the building growing. It's hungry, now that you're here. It wants to digest you to fuel its growth. To reach the sky and grab hold of the last piece of sunlight." The creature walked into the living room, stretching its neck upward. "Look at the light. It's fading."
I looked up at the ceiling light, a single bare bulb that had hung there since I moved in. It had not dimmed noticeably in the three years I had lived there.
"It's not the bulb," I said. "It's the battery."
"There is no battery," the creature said. "The light is powered by the life force of the ten of you who still live in this building. Every time one of you dies, the light dims. But none of you have died yet, so it stays bright." The creature circled me, inspecting me as if I were a new object of interest. "You are the tenth. Roy Baty knows; that's why he was so eager to get you to undergo fusion. Because when you die, he doesn't have to die with you. He can live on without the light. But you're different. You're the anchor."
"Roy Baty is an android," I said, remembering what the creature had said earlier. "So he doesn't feel fear."
"He doesn't feel anything," the creature agreed. "Except hunger."
I walked into the living room, onto the familiar couch. I sat down and put my feet up, trying to calm myself. "If I undergo fusion," I said, "I lose my memories. I become something else."
"That's what you're doing now," the creature said. "Sitting here with your feet up. You're already fusing. You've been for weeks. That's why you think you're pregnant; you're carrying the building in your stomach."
I stood up, walking into the kitchen to get a glass of water. The tap ran clear, cold water, and I filled the glass. "Do you know how long I've been here?" I asked the creature.
"We've always been here," it said. "Since the last human died. We've been waiting for you. The androids came first, of course. They arrived on a ship from the colonies, and then they found you, sitting on your porch in your subdivision, in your town house that was abandoned. They brought you here, to the edge of the world, where the dust doesn't reach. Where the silence is absolute."
I drank the water and felt the cool liquid settle in my stomach. "I don't feel any hunger."
"You're not hungry," the creature said. "You're full. You're full of the sky."
I set the glass down and walked back into the living room. "I need a weapon."
The creature gestured toward the hall closet. "Take a lamp."
I opened the closet door and dragged out a long, thin lamp with a heavy shade. "I need something to cut the cord," I said.
"Use the scissors," the creature said.
I rummaged in the bathroom cabinet and found a pair of cuticle scissors. I returned to the living room and stood facing the creature, the scissors in my hand.
"Are you sure?" I asked. "About the light?"
The creature smiled, showing many small, sharp teeth. "You're going to die. The light is just going to outlive you."
I held the scissors up to the electric cord, which ran along the baseboard of the wall. "If I cut the cord," I said, "I go into total darkness."
"That's right," the creature said. "You'll have to use your phone to call for help. But the phone is also electric."
"I'll charge it," I said. I walked to the bedroom to get the phone.
"You're not an animal," the creature called after me. "You don't need to recharge your battery."
I paused, took a breath, and then continued to the bedroom. The creature followed, its footsteps echoing.
"Why are you doing this?" I asked it as I picked up the receiver. "If I dial, maybe someone will talk to me."
The creature stood on the bed, spreading its legs to reach across the bed to the lamp. "I don't know. I'm doing this because I want to. Because I'm bored. There are no more humans to kill, no more androids to retire. Just you, sitting there on the phone, waiting for a response that will never come."
I dialed my building manager's number. "Hello," I said when he answered. "This is Diana."
"Yes, I know you," he said. "Apartment 2104. How are you?"
"I'm fine," I said. "Did you hear about the fusion tonight? Are you sure about the timing?"
"The timing is perfect," he said. "You're sure you're ready to undergo it?"
"I'm sure," I said. "I'm just curious as to why they're pushing it. Why not wait?"
"Because of the population density," he said. "You know that. They're running out of room."
"Is that why they're telling me it's time?" I asked. "Because they're afraid I won't go through with it? If I don't, they'll have to assign someone else."
"Go through with it," Mr. Kintner said. "And then you can call me back." He hung up.
I hung up, too, and returned to the living room where the creature waited. "Okay," I said to it. "I'll call you back. See if I can get a doctor to come down here."
"You won't be able to reach a doctor," the creature said. "The last of them are on the Moon."
I walked over to it, facing it directly. "What is your name?" I asked.
The creature shrugged. "I don't have a name. I'm a manifestation of the void."
"Do you feel pain?" I asked.
"No."
"Do you feel pleasure?"
"Not really. More of an absence." The creature reached out and brushed a stray hair from my cheek. "You look tense. Relax. Everything will be alright. You've always known everything would be alright, even when it wasn't."
I opened my mouth to say something but then the phone rang.
"Hello," I said into the receiver, my hand shaking.
"Miss Diana?" a woman's voice said. "This is Mrs. Gilt from the fusion center. Are you ready to undergo fusion? We're going to have to start the process; the arc of the Van Allen radiation is very tight tonight and we only have a limited time."
"Yes," I said. "I'm ready."
"Good. We'll have a representative of the fusion council come down to your apartment in about an hour. Until then try to rest. Don't overthink it. Just sit and think about how much you love your life. That emotional energy will help your android brain cells fuse together."
"Thank you," I said, and hung up. "See?" I said to the creature. "They want me to feel love. Like I'm human."
The creature said, "You're not human. You're a special, like Roy Baty said. A human special."
"I'll go to bed," I said. "I'll forget about it and then go back to the phone." I started up the stairs, my steps echoing.
"You're not going to bed," the creature said, following me. "You're going to sit and watch TV until your eyes bleed. Then you'll go to bed, and then you'll get up and make dinner."
"I'm not hungry," I said.
"You are," the creature said. "You're starving."
I stood still, letting the creature taunt me. "Why are you doing this?" I asked again.
"Because I'm lonely," the creature said.
I turned to face it. "You're not lonely. You're everywhere."
"But you're not here," the creature said. "You're up here in your head, in this abstract space you call your mind. I can't reach you." The creature extended its hand and I saw that it had become translucent, like a mist. "I'm going to have to go. I have other lives to live. I have to be the wind, the rain, the decay that kills the last plant on Earth."
"You're part of the building," I said. "You're the dust."
"The dust is alive," the creature said. "That's why the building is growing. It's not sinking; it's blooming."
I felt dizzy. "It's not sinking," I said. "It's growing into a flower."
"But flowers die," the creature said.
I walked into the bedroom, grabbed a pen, and a pad of paper. "I'll write down the exact time the first petal falls," I said. "And then I'll know when the building explodes."
"Use the phone," the creature said. "Call me."
I walked back into the living room, picked up the phone receiver again. "Who am I calling?" I asked the creature.
"The fusion council," it said. "To ask them if you're really a special. Ask them if you have a soul."
"Wouldn't that end the fusion?" I asked. "If I found out I had a soul, I couldn't fuse. That's the whole basis; you have to believe you're an android. That's what Roy Baty said; he said I looked like an android."
"But you aren't," the creature said. "You're a human woman who has been living in a building that's alive for three years. You've started to merge. That's why you're hallucinating. That's why you think you're pregnant."
I hung up the phone. "I'm not pregnant," I said firmly. "I'm not carrying a synthetic child. I'm carrying the spider."
"Which spider?" the creature asked.
"The one I found when I moved in," I said. "The one that hatched in the kitchen drawer. The one that crawled all over me while I was sleeping."
"Did you kill it?"
"No." I had never seen it again after it hatched. Maybe it had gone elsewhere in the apartment. Or maybe it had died. "It's probably still alive," I said. "Hiding in the ceiling."
"The spider is the spider," the creature said. "It doesn't care about you. It just wants to lay its eggs and then die. That's what the building wants, too. To lay its eggs and then die."
I stood for a long time, the phone hanging in my hand. "Okay," I said finally. "I'll call them. I'll ask if I have a soul."
"And if you do," the creature said, "you'll give up on fusion."
"I'll see." I dialed the fusion council's operator. "Hello," I said. "Is Miss Marsten there?" I waited. "Yes, I'm told she's busy." I hung up. "I'll try again later."
The creature said, "You're going to die anyway. Even if the building doesn't crush you." It uncoiled itself and settled onto the couch, curling up like a cat. "I'll just sit here and watch you die."
I walked to the couch and sat down beside it. The creature was warm. "Will you tell me a story?" I asked it. "Before I hang up the phone?"
The creature sighed, its breath ruffling the surface of my hand. "I never heard of a story. It's when you tell a story to an audience. Like when you tell it to a rock."
"I want to tell it to you," I said.
"Then I'm not listening," the creature said.
I smiled. "Goodnight." I stood up and walked into the kitchen to get a bowl of milk.
