Instructions to use psikosen/mini-swarm-browser-258m with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use psikosen/mini-swarm-browser-258m with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="psikosen/mini-swarm-browser-258m", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("psikosen/mini-swarm-browser-258m", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use psikosen/mini-swarm-browser-258m with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "psikosen/mini-swarm-browser-258m"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "psikosen/mini-swarm-browser-258m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/psikosen/mini-swarm-browser-258m

SGLang

How to use psikosen/mini-swarm-browser-258m with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "psikosen/mini-swarm-browser-258m" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "psikosen/mini-swarm-browser-258m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "psikosen/mini-swarm-browser-258m" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "psikosen/mini-swarm-browser-258m",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use psikosen/mini-swarm-browser-258m with Docker Model Runner:
```
docker model run hf.co/psikosen/mini-swarm-browser-258m
```

🐝 Mini-Swarm-Browser (258M)

Mini-Swarm-Browser is an ultra-fast, cooperative multi-agent browser navigation model and architecture based on Canopy-R3 258M.

Instead of routing massive full-DOM trees and screenshots to a monolithic 14B–70B model, Mini-Swarm employs a collaborative swarm of specialized sub-agents that operate concurrently on distilled accessibility trees (AXTree) and Set-of-Marks tags (@e1, @e2), completing real browser actions in under 5 seconds on consumer hardware.

⚡ Key Highlights & Benchmark Stats

Total Parameter Count: 258,555,654 (~258.5M parameters)
Active Parameters / Token: ~112M parameters
Memory Footprint: ~566 MB weights (FP16/BF16) / < 1.2 GB active VRAM
Interaction Speed: ~5 seconds for end-to-end multi-field form completion
Token Efficiency: ~3,500+ tokens saved per action step via Set-of-Marks and AXTree distillation (reducing context from 40k+ tokens to ~400 tokens)
Cooperative Swarm Architecture:
1. Coordinator: High-level semantic planning and goal decomposition.
2. DOM Distiller: AXTree pruning, Set-of-Marks interactive visual tagging.
3. Speculative Action Engine: Multi-action batch emission (predicting dependent steps like click-then-type in a single pass).
4. Fast Verifier: Real-time outcome assertions and local error diagnosis.

🚀 Quickstart

1. Load the Model with Hugging Face Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "psikosen/mini-swarm-browser-258m"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

prompt = "<|im_start|>user\nGoal: Click search button and type 'AI agents'\nElements: [@e1: search_input, @e2: search_btn]<|im_end|>\n<|im_start|>assistant\n"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=64)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

2. Run the Swarm Live Browser Controller

Clone the repo and run the interactive browser agent:

git clone https://huggingface.co/psikosen/mini-swarm-browser-258m
cd mini-swarm-browser-258m

# Launch live headed browser test
python test_interactive.py --workflow form

🛠️ Multi-Agent Architecture

User Objective
       │
       ▼
┌──────────────────────────────┐
│     Swarm Coordinator        │
└──────────────┬───────────────┘
               │
       AXTree / SoM Tokens
               │
       ┌───────┴───────────────┐
       ▼                       ▼
┌──────────────┐        ┌──────────────────────────────┐
│ DOM Distiller│        │  Speculative Action Engine   │
└──────────────┘        │   (Canopy-R3 258M Planner)   │
                        └──────────────┬───────────────┘
                                       │ Action Batch: [@e1, @e2]
                                       ▼
                        ┌──────────────────────────────┐
                        │      Fast Verifier HUD       │
                        └──────────────────────────────┘

📜 License

Released under the Apache 2.0 License.

Downloads last month: -

Safetensors

Model size

0.3B params

Tensor type

BF16