
Qwen 3.5 is Alibaba's latest model family, featuring a hybrid architecture that combines Gated DeltaNet (linear attention) with standard attention in a 3:1 ratio. This design enables dramatically faster inference — up to 8.6x at 32K context and 19x at 256K — while maintaining strong reasoning capabilities. The medium models released February 24, 2026 include:

<table>
  <thead>
    <tr>
      <th>Model</th>
      <th>Total Params</th>
      <th>Active Params</th>
      <th>Architecture</th>
      <th>VRAM (BF16)</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Qwen3.5-122B-A10B</td>
      <td>122B</td>
      <td>10B</td>
      <td>MoE</td>
      <td>~244 GB</td>
    </tr>
    <tr>
      <td>Qwen3.5-35B-A3B</td>
      <td>35B</td>
      <td>3B</td>
      <td>MoE</td>
      <td>~66 GB</td>
    </tr>
    <tr>
      <td>Qwen3.5-27B</td>
      <td>27B</td>
      <td>27B</td>
      <td>Dense</td>
      <td>~54 GB</td>
    </tr>
  </tbody>
</table>

The MoE models route each token through 8 of 256 experts plus 1 shared expert, so only a fraction of the total parameters are active per token. The 35B-A3B model activates just 3B parameters per token while having 35B total — giving you large-model quality at small-model inference cost. All three models are Apache 2.0 licensed, so no HuggingFace token is needed.

In this guide, we'll deploy Qwen3.5-35B-A3B on Vast.ai using SGLang, which has day-0 Qwen3.5 support as of v0.5.9, on a single A100 80GB GPU.

## Deploying on Vast.ai

Install the Vast.ai CLI and set your API key:

```bash
pip install --upgrade vastai
export VAST_API_KEY="YOUR_KEY_HERE"
vastai set api-key $VAST_API_KEY
```

Search for a suitable GPU. The model needs ~66 GB VRAM in BF16, so an 80 GB card works:

```bash
vastai search offers "gpu_ram>=80 num_gpus=1 direct_port_count>=1 rentable=true disk_space>=200 cuda_vers>=12.2" -o dph
```

Deploy the model using SGLang's Docker image. Replace `YOUR_INSTANCE_ID` with an offer ID from the search results:

```bash
vastai create instance YOUR_INSTANCE_ID \
    --image lmsysorg/sglang:latest \
    --env '-p 8000:8000' \
    --disk 200 \
    --onstart-cmd "python3 -m sglang.launch_server \
        --model-path Qwen/Qwen3.5-35B-A3B \
        --host 0.0.0.0 --port 8000 \
        --tp-size 1 \
        --context-length 32768 \
        --reasoning-parser qwen3 \
        --mem-fraction-static 0.85"
```

Key flags:
- `--reasoning-parser qwen3` enables the model's built-in thinking mode, separating reasoning from the final answer
- `--mem-fraction-static 0.85` allocates 85% of GPU memory for KV cache
- `--context-length 32768` sets a 32K context window (the model supports up to 262K natively)

The model will download (~70 GB) and load automatically. Check `vastai logs YOUR_INSTANCE_ID` — you'll see "The server is fired up and ready to roll!" when it's serving.

## Calling the Model

Find your instance's IP and port in the [Instances tab](https://cloud.vast.ai/instances/) of the Vast.ai console. Click the IP address button to see the port mapping for 8000/tcp.

Install the OpenAI SDK:

```bash
pip install --upgrade openai
```

Then call the model:

```python
from openai import OpenAI

VAST_IP_ADDRESS = "YOUR_IP"
VAST_PORT = "YOUR_PORT"

client = OpenAI(
    api_key="EMPTY",
    base_url=f"http://{VAST_IP_ADDRESS}:{VAST_PORT}/v1"
)

response = client.chat.completions.create(
    model="Qwen/Qwen3.5-35B-A3B",
    messages=[
        {"role": "user", "content": "What are three benefits of mixture-of-experts models? Be concise."}
    ],
    max_tokens=1024,
    temperature=0.7
)

print(response.choices[0].message.content)
```

### Response

```
1. **Computational Efficiency:** Only a subset of parameters is activated per input,
   reducing inference costs while maintaining a massive total parameter count.
2. **Enhanced Performance:** The larger effective model capacity allows for better
   accuracy on complex and diverse tasks compared to dense models of similar compute.
3. **Specialization:** Individual experts can learn to handle specific input patterns
   or domains, improving the model's robustness and adaptability.
```

With `--reasoning-parser qwen3`, the model's thinking process is automatically separated from the final answer. The thinking tokens appear in `reasoning_content` in the API response, while the clean answer appears in `content`.

## Other Models

You can swap `Qwen/Qwen3.5-35B-A3B` for the other medium models:

- **Qwen3.5-122B-A10B**: Larger MoE model, 122B total / 10B active. At BF16 the full weights need ~244 GB, so you'll need multiple GPUs (e.g., 4× A100 80GB with `--tp-size 4`). Alternatively, GGUF quantizations are available — the Q4_K_M variant is ~70 GB and fits on a single 80GB card.
- **Qwen3.5-27B**: Dense model, all 27B parameters active. Needs ~54 GB VRAM. Fits on a single A100 or H100.

## Conclusion

Qwen3.5-35B-A3B packs 35B parameters into a model that only activates 3B per token, and its hybrid DeltaNet architecture delivers substantially faster inference than standard transformers. With SGLang and a single A100 on Vast.ai, you can have it serving in under 10 minutes.

For production use, consider increasing `--context-length` — the model supports up to 262K natively, with 128K+ recommended for best reasoning performance. GGUF quantizations (Q4_K_M at ~22 GB) are also available for running on consumer GPUs like the RTX 4090.

## Useful Resources

- [Qwen3.5-35B-A3B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-35B-A3B)
- [Qwen3.5-122B-A10B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-122B-A10B)
- [Qwen3.5-27B on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-27B)
- [SGLang Documentation](https://docs.sglang.ai/)
- [Vast.ai Documentation](https://docs.vast.ai/)


Deploy Qwen 3.5 medium models on Vast.ai using SGLang, including setup, launch commands, and API usage.

Running Qwen 3.5 Medium Models on Vast.ai


Liquid AI just released [LFM2-24B-A2B](https://huggingface.co/LiquidAI/LFM2-24B-A2B), a 24 billion parameter foundation model that only activates 2.3 billion parameters per token. That ratio — less than 10% of total parameters active at inference time — makes it one of the most aggressively sparse mixture-of-experts models available today.

What makes LFM2 different from other MoE models isn't just the sparsity. The architecture is a hybrid design that replaces most of the attention layers with gated short convolution blocks. Out of 40 total layers, only 10 use grouped query attention — the other 30 are convolutional. This is a fundamentally different approach from transformer-only models, and it's what allows Liquid AI to push the active parameter count so low while maintaining competitive quality across benchmarks like MMLU-Pro, GPQA Diamond, IFEval, and MATH-500.

Liquid AI, an MIT spinoff, has scaled this architecture from 350M to 24B parameters following log-linear scaling laws. LFM2 is the first model in the family large enough to be broadly useful — and with day-one support for vLLM, SGLang, and llama.cpp, it's ready to deploy today.

> **Note**: LFM2-24B-A2B is an early checkpoint trained on 17 trillion tokens without reinforcement learning or extensive post-training. A post-trained version (LFM2.5-24B-A2B) is coming.

## Model Overview

<table>
  <thead>
    <tr>
      <th>Property</th>
      <th>Value</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>Developer</td>
      <td><a href="https://www.liquid.ai">Liquid AI</a></td>
    </tr>
    <tr>
      <td>Model</td>
      <td>LFM2-24B-A2B</td>
    </tr>
    <tr>
      <td>Architecture</td>
      <td>Hybrid MoE: 30 convolutional + 10 GQA attention layers, 64 experts, top-4 routing</td>
    </tr>
    <tr>
      <td>Total Parameters</td>
      <td>24B</td>
    </tr>
    <tr>
      <td>Active Parameters</td>
      <td>2.3B per token</td>
    </tr>
    <tr>
      <td>Context Length</td>
      <td>32,768 tokens</td>
    </tr>
    <tr>
      <td>Training Data</td>
      <td>17T tokens (mixed BF16/FP8)</td>
    </tr>
    <tr>
      <td>License</td>
      <td><a href="https://www.liquid.ai/lfm-open-license-v1-0">LFM Open License v1.0</a></td>
    </tr>
    <tr>
      <td>HuggingFace</td>
      <td><a href="https://huggingface.co/LiquidAI/LFM2-24B-A2B">LiquidAI/LFM2-24B-A2B</a></td>
    </tr>
  </tbody>
</table>

## Deploy LFM2-24B-A2B on Vast.ai with vLLM

At BF16, the model weights are ~48 GB — it fits on a single A100-80GB with room for a 32K context KV cache. No quantization or multi-GPU setup needed.

### Prerequisites

- A [Vast.ai](https://cloud.vast.ai) account with credits
- The Vast CLI installed and configured:

```bash
pip install --upgrade vastai  # make sure you're on v0.5.0+
vastai set api-key <YOUR_API_KEY>
```

### Find a GPU

Search for a single GPU with at least 80 GB VRAM and direct port access:

```bash
vastai search offers "gpu_ram>=80 num_gpus=1 direct_port_count>=1 cuda_vers>=12.4 cuda_vers<13" -o dph
```

> **Important**: Use a machine with **CUDA 12.x** (not 13.x). The vLLM nightly image is built against CUDA 12 and will fail with a driver mismatch on CUDA 13 machines.

### Launch the Instance

Pick an instance ID from the search results and deploy with vLLM. LFM2 requires a newer version of the `transformers` library than what ships in the vLLM image, so we upgrade it at startup:

```bash
vastai create instance <INSTANCE_ID> \
  --image vllm/vllm-openai:nightly \
  --env "-p 8000:8000" \
  --disk 100 \
  --onstart-cmd "pip install 'transformers>=5.0.0' && \
    vllm serve LiquidAI/LFM2-24B-A2B \
    --host 0.0.0.0 --port 8000 \
    --max-model-len 32768 --dtype auto \
    --trust-remote-code"
```

This upgrades the tokenizer library, pulls the model from HuggingFace, and starts an OpenAI-compatible API server on port 8000. Initial startup takes several minutes while the model downloads (~48 GB for the BF16 weights).

Once the instance is running, find your IP and port in the [Vast console](https://cloud.vast.ai/instances/) — click the IP button on your instance to see the port mapping.

> **Why `nightly` and `transformers>=5`?** LFM2's tokenizer uses the `TokenizersBackend` class introduced in `transformers 5.0`. The stable vLLM image pins `transformers<5`, so the tokenizer fails to load. Using `nightly` with a transformers upgrade resolves this. This was tested with vLLM nightly build `v0.16.0rc2`. Once a stable vLLM release ships with `transformers>=5.0.0` support, you can switch to `vllm/vllm-openai:latest` and drop the `pip install` step.

## Call the API

```bash
curl http://<VAST_IP>:<PORT>/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "LiquidAI/LFM2-24B-A2B",
    "messages": [{"role": "user", "content": "What makes mixture-of-experts models efficient?"}],
    "max_tokens": 300
  }'
```

## Cleanup

When you're done, destroy the instance to stop charges:

```bash
vastai destroy instance <INSTANCE_ID>
```

## Resources

- [LFM2-24B-A2B on HuggingFace](https://huggingface.co/LiquidAI/LFM2-24B-A2B)
- [LFM2-24B-A2B GGUF Quantizations](https://huggingface.co/LiquidAI/LFM2-24B-A2B-GGUF)
- [Liquid AI Blog: LFM2-24B-A2B](https://www.liquid.ai/blog/lfm2-24b-a2b)
- [Liquid AI Models](https://www.liquid.ai/models)
- [Vast.ai Documentation](https://docs.vast.ai/)
- [vLLM Documentation](https://docs.vllm.ai/)


Liquid AI released LFM2-24B-A2B, a sparse 24B MoE model with 2.3B active parameters per token.

Posts about: Moe

Running Qwen 3.5 Medium Models on Vast.ai

Liquid AI's LFM2 Just Dropped — Here's How to Run It on Vast.ai