selfhost.directory

The index of self-hostable projects — ranked, categorized and served with live updates.
Own your data and stop giving it to big tech.

inclusionAI logo

ZwZ2-9B

inclusionAI·Instruct·released 16 Oct 2026

Unknown licence9.4B256K ctx ≈7.1 GB
Your GPU
llama-server -hf inclusionAI/ZwZ2-9B-GGUF:Q4_K_M
All 4 runners ↓

ZwZ2-9B is the Qwen3.5-9B-based model in Zooming-without-Zooming-v2 (ZwZ2). It learns to connect visual evidence across regions through comparison, aggregation, matching, and multi-step reasoning. Training uses full images; inference requires no zooming tools.

Notify me when

How much VRAM does ZwZ2-9B need?

At Q4_K_M ZwZ2-9B needs about 7.1 GB with an 8K context (7.1 GB at Q4, 20 GB at full precision). Pick your GPU to see what fits and how fast it runs.

--vram
any
--gpus
--quant

Needs: ≈7.0 GB · Q4_K_M · 8K context — Pick a GPU to check the fit.

Change quant, context or KV cache to compare
Advanced: context 8K · KV cache 16-bit
Context
KV cache

Smallest GPU for each quant and context (KV cache 16-bit)

Quantisation8K16K32K64K128K256K
Q3_K_M
Q4_K_M
Q5_K_M
Q6_K
Q8_0
FP16

≤16 GB card24–32 GB48–96 GBmore

QuantisationQualityWeightsKV cacheTotal
Q2_KHeavy lossSmallest, noticeably worse qualityHeavy loss2.9 GB est.1.0 GB≈4.5 GB
Q3_K_MNoticeable lossSmall, some quality lossNoticeable loss4.3 GB file1.0 GB≈6.0 GB
Q4_K_MrecommendedSmall lossRecommended default for local useSmall loss5.2 GB file1.0 GB≈7.0 GB
Q5_K_MNear-losslessNear-lossless for most tasksNear-lossless6.0 GB file1.0 GB≈7.8 GB
Q6_KNear-losslessPractically losslessNear-lossless6.8 GB file1.0 GB≈8.7 GB
Q8_0LosslessLossless in practice, twice the size of Q4Lossless8.9 GB file1.0 GB≈11 GB
FP16 / BF16LosslessFull precision, as releasedLossless17 GB file1.0 GB≈19 GB

Estimates for 8K tokens of context: weights + KV cache + runtime overhead. Speed assumes memory-bound decoding at your GPU’s bandwidth. Pick a GPU above and the fit column checks every row against it. Weights marked “file” use the real GGUF download size; “est.” rows are computed from the parameter count.

How to pick a quant

Take the biggest quant that fits with room left for context. Q4_K_M is the usual sweet spot. Below Q4 quality drops fast, above Q6 you mostly pay memory for nothing. Big models (70B+, large MoE) tolerate low quants better than small ones.

  • LosslessF16, Q8_0
  • Near-losslessQ6_K, Q5_K_M
  • Small lossQ5_K_S, Q4_K_M, IQ4_NL, IQ4_XS, Q4_K_S
  • Noticeable lossQ4_0, Q3_K_L, Q3_K_M, IQ3_M
  • Heavy lossIQ3_XXS, Q3_K_S, IQ2_M, Q2_K

Typical quality loss, not measured per model. Tiers come from KL divergence against FP16 on Llama 3 8B: llama.cpp perplexity README: LLaMA 3 8B scoreboard (KLD and ΔPPL vs FP16, Wikitext-2); llama.cpp quantize tool: ΔPPL per quant type @ Llama-3-8B

How much context fits?

Model quant
KV cache

At 8K context, Q4_K_M needs ≈7.0 GB(5.2 GB weights + 1.0 GB KV cache + 0.8 GB overhead). Pick a GPU above to see how much context fits.

0102030401K2K4K8K16K32K64K128K256KGB

Weights + overheadKV cache

Can the RTX 4060 run ZwZ2-9B? →

Speed vs quality per quant

Smaller quants run faster and fit in less memory, but lose quality.

Q2_KQ3_K_MQ4_K_MQ5_K_MQ6_KQ8_0F16

Fits in VRAMTight fitCPU offload✕ won't runQuality is typical KL divergence vs FP16, not measured for this model.

Real-world speed of ZwZ2-9B

Tokens per second measured by people running it at home. A GPU shows a median once 3 reports agree on the same setup. Reports far from our estimate wait for a review. Measure yours with llama-bench →

No medians yet. Be one of the first 3 to report a setup.

All reports

Run ZwZ2-9B locally

--quantQ4_K_Mdefault; pick your GPU to tailor itinclusionAI/ZwZ2-9B-GGUF

OpenAI-compatible server with a web UIdocs
llama-server -hf inclusionAI/ZwZ2-9B-GGUF:Q4_K_M

Downloads the GGUF on first run. Installed through llama.app? The same flags work with `llama serve`.

Install llama.cpp

Serve an OpenAI-compatible APIdocs
vllm serve inclusionAI/ZwZ2-9B
Docker Compose (NVIDIA)docs
yaml
services:
  vllm:
    image: vllm/vllm-openai:latest
    command: inclusionAI/ZwZ2-9B
    ports:
      - "8000:8000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 8000.

Install vLLM

Serve an OpenAI-compatible APIdocs
sglang serve inclusionAI/ZwZ2-9B --host 0.0.0.0 --port 30000
Docker Compose (NVIDIA)docs
yaml
services:
  sglang:
    image: lmsysorg/sglang:latest
    command: sglang serve inclusionAI/ZwZ2-9B --host 0.0.0.0 --port 30000
    ports:
      - "30000:30000"
    ipc: host
    volumes:
      - ~/.cache/huggingface:/root/.cache/huggingface
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]

API on port 30000.

Install SGLang

Ask about an imagedocs
python
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="inclusionAI/ZwZ2-9B", dtype="auto", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
            {"type": "text", "text": "Describe this image."},
        ],
    }
]
out = pipe(text=messages, max_new_tokens=256)
print(out[0]["generated_text"][-1]["content"])

Install Transformers

Quantised versions and downloads

About ZwZ2-9B

ZwZ2-9B is the Qwen3.5-9B-based model in Zooming-without-Zooming-v2 (ZwZ2). It learns to connect visual evidence across regions through comparison, aggregation, matching, and multi-step reasoning. Training uses full images; inference requires no zooming tools. On 25 benchmarks across six domains, it improves over its base model on every benchmark, increasing overall accuracy from 51.63 to 56.59 (+4.96 points). Gains extend from fine-grained perception to spatial reasoning, long-context understanding, multimodal agents, and general visual reasoning. EvidenceConnect builds verified multi-region questions from SA-1B images: • Group 3–7 related regions and generate questions using privileged zoomed views. • Check answers through cross-model agreement and drop-one region tests. • Convert verified questions to four-option tasks and train with GRPO, without an additional supervised fine-tuning stage. All three model sizes use approximately 20K training examples. The policy sees full images with colored region outlines; privileged crop layouts are used only for data construction. See ZwZ2-RL-VQA for the data schema and synthesis process. Editable vector PDF. The figure shows eight representative benchmarks; the complete results follow.

Full model card on Hugging Face →

Other ZwZ2 models

Whole family →
inclusionAI logo
ZwZ2-4BinclusionAI
4.5BCheck licence≈4.2 GB↓ 1.1K
inclusionAI logo
ZwZ2-27BinclusionAI
27.4BCheck licence≈19 GB↓ 1.5K

Alternatives to ZwZ2-9B

All alternatives →
Qwen logo

Qwen3.5-9B

Qwen·Vision

≈7.2 GB

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.

9.7B256K ctx Apache 2.0
6.2MFeb 2026ollama
Qwen logo≈6.8 GB

Meet Qwen3-VL — the most powerful vision-language model in the Qwen series to date.

8.8B256K ctx Apache 2.0
6.1MOct 2025ollama
Zyphra logo

ZAYA1-VL-8B

Zyphra·Vision

≈6.6 GB

ZAYA1-VL-8B is a vision-language model (VLM) built upon Zyphra's ZAYA1-8B LLM. It has state-of-the-art performance among VLMs for its size and inference efficiency.

9.7B32K ctx Apache 2.0
339.4KApr 2026gguf
Qwen logo

Qwen3.5-9B-Base

Qwen·Vision

≈7.2 GB

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance.

9.7B256K ctx Apache 2.0
266.8KFeb 2026gguf
XiaomiMiMo logo

MiMo-V2.6-Distill-Qwen-9B

XiaomiMiMo·Vision

≈7.1 GB

MiMo-V2.6-Distill-Qwen-9B is a 9B agentic model developed by Xiaomi MiMo through supervised fine-tuning of Qwen3.5-9B on MiMo-generated data. It covers coding, general-purpose agent tasks, visual coding, and cybersecurity.

9.4B256K ctx MIT
38.9KSept 2026gguf
black-forest-labs logo

flux-3-image-klein-pup

black-forest-labs·Vision

≈7.1 GB

FLUX 3 Image Klein Prompt Upsampler is a 9B vision-language model that expands a short image request into the detailed prompt FLUX 3 Image Klein works best with. It takes the request, up to 10 reference images and an optional list of bounding boxes, and returns one prompt.

9.4B256K ctx Apache 2.0
31KOct 2026gguf

ZwZ2-9B FAQ

How much VRAM does ZwZ2-9B need?
About 7.1 GB at Q4_K_M with an 8K context. Q8_0 needs about 11 GB and full precision about 20 GB. Longer contexts need more memory for the KV cache.
Can ZwZ2-9B run on the RTX 4060?
Yes. It fits at Q4_K_M (about 7.0 GB), at an estimated 26 tokens/s. It fits comfortably on the Apple M4 (16 GB), Radeon RX 6700 XT, RTX 3060 12GB.
Can I use ZwZ2-9B commercially?
Unclear. It is released under Unknown licence. No licence stated: assume all rights reserved. Check the licence on the model card before using it in a product.
What is the context length of ZwZ2-9B?
262,144 tokens (256K). The full window needs far more memory than an 8K context.
How do I run ZwZ2-9B with llama.cpp?
Run llama-server -hf inclusionAI/ZwZ2-9B-GGUF:Q4_K_M. It downloads the GGUF and starts an OpenAI-compatible server.

Data from Hugging Face, the model card and public leaderboards. Memory and speed figures are estimates; real numbers depend on the runtime, context length and drivers. Always check the licence yourself before commercial use.