BurrowInstall guideMy licence
The model

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. It runs on your computer; mail and calendar connections go directly to your providers.

Burrow 2B. Your data stays underground. Local AI for your everyday work.
Download Q4_K_M · 1.56 GBView on Hugging Face

Opens Hugging Face. No account is needed. The app and this model are separate downloads; the model runs locally after setup.

Details · How to run it · Scores · Known weaknesses · CPU and integrated GPUs

Use it in Burrow

Burrow is a local desktop assistant for your mail, calendar, notes and To-Do’s. Burrow 2B is its small, purpose-built model for everyday work. Chat is free; mail, calendar and other tools unlock with a Burrow licence. See the installation guide to get started on Windows.

Model details

Base modelopenbmb/MiniCPM5-2B, Apache-2.0 (revision f97400052a43d642bbc6e9975e2397e3ae6a6b52)
MethodLoRA (r=64, alpha 128, all linear layers), supervised fine-tuning, 2 epochs / 1,162 steps; the step-1162 adapter was merged and quantized
Training data7,255 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.
Filesburrow-2b-v1.Q4_K_M.gguf (1,561,319,520 bytes, primary). imatrix from 40 chunks of training text.
SHA-256 (Q4_K_M)9e5cd3bafe8bae016b91b06c2c55c2cb698080d257e8933084a4abbbf40a03e4
Chat templateEmbedded in the GGUF (patched for list-valued tool arguments). No --chat-template-file is needed.
ContextTrained and tested at 32,768 tokens
LicenceThe 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, and the model is only expected to work with them. It is not a general chat model.

How to run it

Download the Q4_K_M file 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 model folder is %LOCALAPPDATA%\Burrow\models , unless overridden. Inference runs locally after downloading; no Hugging Face account is required.

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.

SuiteStock MiniCPM5-2BBurrow 2B v1Bonsai 2 27B (reference)
A, 250 tests (mail, calendar, To-Do’s, 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 runs24 (48.0%), 1 run33, 32, 35 (mean 33.3, 66.7%)38 (76.0%), 1 run
C-dev, 50 hard cases (mined, diagnostic only)254139

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 test ids and prompts are not published.

Known weaknesses

Integrated GPUs and CPU-only laptops

Burrow picks the llama.cpp build for this model with Settings → Models → Server options → Backend: Auto | CPU | GPU.

On CPU (4 threads, development PC with the GPU hidden, Q4_K_M) it decodes about 19 tokens/s and processes about 100 tokens/s of prompt at 3,300 tokens; a 40-test subset of A passed 38 of 40 with a median test latency of 26.7 s.

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):

BackendFirst turnDecode (tok/s)Prompt (tok/s)
CPU86.3 s25.098
Vulkan, -ngl 026.0 s24.2186
Vulkan, -ngl 9916.4 s23.7268

The SYCL build found no device. Answer quality was not measured on the laptop; 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.

Download Q4_K_M · 1.56 GB

Install guide · My licence · Home