Instructions to use TyPlant/Burrow-2B-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use TyPlant/Burrow-2B-GGUF with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible server with a web UI: llama serve -hf TyPlant/Burrow-2B-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf TyPlant/Burrow-2B-GGUF:Q4_K_M
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible server with a web UI: llama serve -hf TyPlant/Burrow-2B-GGUF:Q4_K_M # Run inference directly in the terminal: llama cli -hf TyPlant/Burrow-2B-GGUF:Q4_K_M
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible server with a web UI: ./build/bin/llama serve -hf TyPlant/Burrow-2B-GGUF:Q4_K_M # Run inference directly in the terminal: ./build/bin/llama cli -hf TyPlant/Burrow-2B-GGUF:Q4_K_M
Use Docker
docker model run hf.co/TyPlant/Burrow-2B-GGUF:Q4_K_M
- Llama new
- LM Studio
- Jan
- vLLM
How to use TyPlant/Burrow-2B-GGUF with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "TyPlant/Burrow-2B-GGUF" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "TyPlant/Burrow-2B-GGUF", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/TyPlant/Burrow-2B-GGUF:Q4_K_M
- Ollama
How to use TyPlant/Burrow-2B-GGUF with Ollama:
ollama run hf.co/TyPlant/Burrow-2B-GGUF:Q4_K_M
- Unsloth Desktop
- Pi
How to use TyPlant/Burrow-2B-GGUF with Pi:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf TyPlant/Burrow-2B-GGUF:Q4_K_M
Configure the model in Pi
# Install Pi: npm install -g @earendil-works/pi-coding-agent # Add to ~/.pi/agent/models.json: { "providers": { "llama-cpp": { "baseUrl": "http://localhost:9931/v1", "api": "openai-completions", "apiKey": "none", "models": [ { "id": "TyPlant/Burrow-2B-GGUF:Q4_K_M" } ] } } }Run Pi
# Start Pi in your project directory: pi
- Docker Model Runner
How to use TyPlant/Burrow-2B-GGUF with Docker Model Runner:
docker model run hf.co/TyPlant/Burrow-2B-GGUF:Q4_K_M
- Lemonade
How to use TyPlant/Burrow-2B-GGUF with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull TyPlant/Burrow-2B-GGUF:Q4_K_M
Run and chat with the model
lemonade run user.Burrow-2B-GGUF-Q4_K_M
List all available models
lemonade list
- Hermes Agent
How to use TyPlant/Burrow-2B-GGUF with Hermes Agent:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf TyPlant/Burrow-2B-GGUF:Q4_K_M
Configure Hermes
# Install Hermes: curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash hermes setup # Point Hermes at the local server: hermes config set model.provider custom hermes config set model.base_url http://127.0.0.1:9931/v1 hermes config set model.default TyPlant/Burrow-2B-GGUF:Q4_K_M
Run Hermes
hermes
- Atomic Chat
- OpenClaw
How to use TyPlant/Burrow-2B-GGUF with OpenClaw:
Start the llama.cpp server
# Install llama.cpp: brew install llama.cpp # Start a local OpenAI-compatible server: llama serve -hf TyPlant/Burrow-2B-GGUF:Q4_K_M
Configure OpenClaw
# Install OpenClaw: npm install -g openclaw@latest # Register the local server and set it as the default model: openclaw onboard --non-interactive --mode local \ --auth-choice custom-api-key \ --custom-base-url http://127.0.0.1:9931/v1 \ --custom-model-id "TyPlant/Burrow-2B-GGUF:Q4_K_M" \ --custom-provider-id llama-cpp \ --custom-compatibility openai \ --custom-text-input \ --accept-risk \ --skip-health
Run OpenClaw
openclaw agent --local --agent main --message "Hello from Hugging Face"
Meet Burrow ยท Download the app ยท Installation guide ยท Download the model
Put Burrow 2B to work in Burrow
Burrow is a local desktop assistant for your mail, calendar, notes and to-dos. Burrow 2B is its small, purpose-built model for everyday work, including laptops without a dedicated GPU. Your model runs on your computer; mail and calendar connections go directly to your providers.
Get Burrow for Windows or Linux / Chromebook โ Chat is free. Mail, calendar and other tools unlock with a Burrow licence. The app and this 1.56 GB model are separate downloads; see the installation guide to get started.
Burrow 2B v1
A 2.6B-parameter model fine-tuned to run Burrow's mail, calendar and to-do tool loop on an ordinary PC, including laptops without a dedicated GPU.
| Base model | openbmb/MiniCPM5-2B, Apache-2.0 (revision f97400052a43d642bbc6e9975e2397e3ae6a6b52) |
| Method | LoRA (r=64, alpha 128, all linear layers), supervised fine-tuning, 2 epochs / 1,162 steps; the step-1162 adapter was merged and quantized |
| Training data | 7,255 synthetic Burrow trajectories (tool calls, tool errors and corrections, multi-turn threads, thinking and non-thinking turns) written by a teacher model (opencode-go/muse-spark-1.3-contributor) in 22 fixture worlds, plus 2,305 thinking sequences; screened against every evaluation set (n-gram, embedding and entity overlap, 0 remaining). No real mail or user data. |
| Files | burrow-2b-v1.Q4_K_M.gguf (1,561,319,520 bytes, primary), burrow-2b-v1.Q8_0.gguf (2,679,711,840 bytes, higher quality, more memory) and burrow-2b-v1.Q3_K_M.gguf (1,291,421,792 bytes, smallest, lower quality). imatrix from 40 chunks of training text. |
| SHA-256 (Q4_K_M) | 9e5cd3bafe8bae016b91b06c2c55c2cb698080d257e8933084a4abbbf40a03e4 |
| SHA-256 (Q8_0) | 6237649721b89ac93765094e41bc195a7d0b877fc15798df6ef96ef939dc52e3 |
| SHA-256 (Q3_K_M) | 778696d58da405c51e3db55262efd8afbfe4480de4a2b68821b350801e753139 |
| Chat template | embedded in the GGUF (patched for list-valued tool arguments; template sha256 6b32756a84b22bf3347b6c9e47d2b4f7a0141bc908c0b493b1781f475d780975). No --chat-template-file is needed. |
| Context | trained and tested at 32,768 tokens |
| Licence | The base model is Apache-2.0. Burrow 2B is a derivative of it. The training data was produced with a hosted teacher model; check that provider's terms before redistributing the weights. |
Burrow's tool names, tool schemas and system prompt were fixed before training (sha256 10dd5cac... and e4d4d441...) and the model is only expected to work with them. It is not a general chat model.
How to run it
The primary Q4_K_M file, and the Q8_0 and Q3_K_M variants, are available from TyPlant/Burrow-2B-GGUF. The public Burrow app archive does not include model weights. Builds with Hugging Face support offer Settings > Models > Burrow 2B v1 (Q4_K_M) > Download, then Load. Older releases can use the same file: copy it to Burrow's model folder under burrow-2b/, or use Register another GGUF to select it. The installed app's model folder is %LOCALAPPDATA%\Burrow\models on Windows and ~/.local/share/Burrow/models on Linux (unless overridden). Inference runs locally after downloading; no Hugging Face account is required. The in-app download fetches only Q4_K_M; to use Q8_0 or Q3_K_M, download the file and use Register another GGUF. A Q6_K build exists locally but is not published.
By hand, with llama.cpp build 11723 (31288c9c6), the flags it was evaluated with are:
llama-server -m burrow-2b-v1.Q4_K_M.gguf --jinja -fa on -c 32768 -np 1 -ngl 99 -ctk q8_0 -ctv q8_0 --reasoning-budget 256
--cache-ram 2048 --top-k 20 --top-p 0.95 --min-p 0 --host 127.0.0.1 --port 8101
--min-p 0 matters: llama.cpp's default of 0.05 can make the 2B loop. Do not use n-gram speculation (not tested). The GPU evaluations used --reasoning-budget 1024; the CPU and integrated-GPU profiles use 256.
Scores
All on Burrow's own harness (v1.1) on one RTX 4060, Q4_K_M, serial runs; the stock model is MiniCPM5-2B Q8_0 with the same harness. Higher is better.
| suite | stock MiniCPM5-2B | Burrow 2B v1 | Bonsai 2 27B (reference) |
|---|---|---|---|
| A, 250 tests (mail, calendar, to-dos, safety, dates) | 184 (73.6%) | 228 (91.2%) | not run |
| B, 232 tests (a second world, different phrasings) | 133 (57.3%) | 178 (76.7%) | not run |
| C-test, 50 sealed hard cases, 3 runs | 24 (48.0%), 1 run | 33, 32, 35 (mean 33.3, 66.7%) | 38 (76.0%), 1 run |
| C-dev, 50 hard cases (mined, diagnostic only) | 25 | 41 | 39 |
Other quants, C-test only (one run each, same harness; Q4_K_M above is the mean of three): Q8_0 32/50 (64.0%), Q3_K_M 27/50 (54.0%). A, B and C-dev were not run for them. The Q3_K_M gap is larger than the run-to-run spread of about 3 tests; Q8_0 is within it, so Q8_0 is not shown to be better than Q4_K_M.
Read C with its limits. C was built by hand-picking hard cases, so its difficulty is constructed: the gaps between models matter, the absolute scores do not. Of its 100 tests, 79 carry exactly two axes, 65 are single-turn, and three of its axes have fewer than 10 tests per half, so no per-axis conclusion is drawn. A and B were visible during development and are informational, not held-out. The run-to-run spread of v1 on C-test is about 3 tests. Scores depend on the clock and on the stamp format (below); the A/B/C test ids and prompts are not published.
Known weaknesses
- Dates and times. Date and time tests are the weakest area: 23 of 35 (65.7%) across A and C-test.
- Plain text instead of a tool call. On the 342-turn validation set the model answered in plain text on 9 of the 202 turns that needed a tool call (4.5% per turn; 95.5% structured against a 98% gate, which it failed and which was accepted as a known weakness). End to end, over the 72 A-visible and C-dev tests that need a tool, none ended that way, because the agent loop re-prompts. Arguments are valid JSON in 100% of cases and no tool text leaks into replies.
- Sensitivity to the stamp format. Every user turn starts with a
[Now: ...]line. Adding a 14-day day-by-day list to that line (an experiment, since reverted) cost v1 about 6 C-test tests (33.3 to 27.0 mean) without any change to the model, while A moved by 2 tests. A fine-tune on the longer stamp did not recover it. Keep the stamp exactly as the app writes it. - Free-text arguments. About half of the validation tool calls match the teacher's wording exactly; the misses inspected were valid alternatives (a different search term, a different email body).
- Small model limits. No vision add-on, and it was not tested beyond a 32k context.
Integrated GPUs and CPU-only laptops
Burrow picks the llama.cpp build for this model with Settings > Models > Server options > Backend: Auto | CPU | GPU.
- Auto uses the NVIDIA GPU (the CUDA build, exactly the command used for the scores above) if there is one. On a PC whose only GPU is an Intel (Iris Xe, UHD, Arc) or AMD (Radeon) integrated GPU that the Vulkan build can see, it uses the profile named by
default_profile.igpuinassistant-app\burrow\backend_profiles.json. That isvulkan_ngl99(all layers on the iGPU, 4 threads), chosen by the Latitude 5520 benchmark below; on PCs without a usable GPU it is always the CPU profile. If the Vulkan profile fails to start the app falls back to CPU. - CPU forces the CPU build (
tools\llama.cpp-cpu). GPU forces the best GPU profile for the hardware (CUDA on NVIDIA, Vulkan with all layers on the GPU on Intel and AMD). - If a Vulkan or SYCL profile does not start serving within 90 seconds, or its first request errors, Burrow restarts on the CPU profile, writes the reason to
logs\backend.log, and keeps using the CPU until you change the Backend setting. To force the CPU on any PC, set Backend to CPU. - Builds go in
tools\llama.cpp-cpu,tools\llama.cpp-vulkanandtools\llama.cpp-sycl(official llama.cpp releaseb11723, the same commit as the CUDA build). The CPU and iGPU profiles use--reasoning-budget 256.
Speed on CPU (4 threads, measured on the development PC with the GPU hidden, Q4_K_M): about 19 tokens/s decode and about 100 tokens/s prompt processing at 3,300 tokens; a 40-test subset of A passed 38 of 40 with a median test latency of 26.7 s. Those are development-PC numbers. Measured on an Intel Iris Xe laptop (Latitude 5520, i7-1185G7, 31.7 GB, AC power, Best performance, 2026-10-27), first turn of a 3,301-token request (52 tokens generated): CPU 86.3 s, Vulkan -ngl 0 26.0 s, Vulkan -ngl 99 16.4 s; decode (llama-bench tg128) 25.0 / 24.2 / 23.7 tok/s, prompt processing (pp3300) 98 / 186 / 268 tok/s. The SYCL build found no device. Not measured on the laptop: answer quality (the 40-test subset was not run there); on the RTX 4060 the Vulkan build scored 27/50 on C-test against 33/32/35 on CUDA (one run), so set Backend to CPU if answers look worse.
- Downloads last month
- 209
3-bit
4-bit
8-bit
Model tree for TyPlant/Burrow-2B-GGUF
Base model
openbmb/MiniCPM5-2B