vasudha-4b-v3-gguf

Vasudha 4B — v3 (GGUF)

A 4B assistant that runs entirely on your own machine, built for engineering and science questions where the answer is a number you can check.

Made by Chaitanya, an independent developer.

What it is for

Vasudha is designed to be used with a Python sandbox, not on its own. It picks the governing equation and sets the problem up; the arithmetic is executed in a real interpreter and the working is shown. Used that way it is a calculator that explains itself. Used without tools it will state numbers it has not computed, and some of them will be wrong.

That distinction is the whole design, and the numbers below say why.

Measured behaviour

Six numeric engineering problems (cantilever deflection and stress, Reynolds number, RC cutoff, projectile range, wall conduction), ground truth computed in Python, greedy decoding:

The harness matters more than the prompt: asking the model to use a calculator got a tool call one time in six; giving it a real tool schema got six out of six.

Known limitations

Stated plainly because you will hit them.

• Standard formulas can be recalled with the wrong boundary condition. For a cantilever with an end load it has produced PL³/(48EI) — the simply-supported centre-load case — instead of PL³/(3EI). It gets the simply-supported version right, so the coefficient is memorised without the support condition attached to it.
• Magnitude errors of exactly 10× or 1000× appear in both arithmetic and recalled facts. Asked for planetary masses it returned Neptune ten times too heavy while getting the ordering of all eight planets right.
• It will cite sources it did not read. Asked for planetary data it ran three searches that returned no figures at all, wrote a table from memory, and signed it "Source: NASA Solar System data sheet". Any application built on this model should track provenance itself rather than trusting the model's account of it.
• Long replies can be cut off mid tool call, which some servers reject outright. Allow at least 2048 output tokens.

None of these are subtle once you look for them, and all of them are why the tool path is not optional.

Files

SHA-256 of the Q3_K_M build:

0a31e817bf8e7a3f9d1214ac1445648e734b3acd10ce019f02fa12b691a791c9

Check it after downloading — a truncated GGUF does not fail cleanly, it fails confusingly.

Use with Ollama

ollama create vasudha -f Modelfile

The GGUF carries its own Qwen3 chat template, including tool-call support, so structured tool calling works without extra configuration.

Use with llama.cpp

llama-server -m vasudha-leaprunning-4b-v3-Q3_K_M.gguf --jinja -c 16384

--jinja matters: it uses the chat template embedded in the file, which is what makes tool calls parse.

Architecture

Qwen3-4B converted to a hybrid: most layers Gated Linear Attention (O(1) memory in context length) with a minority kept as full attention, plus MoE upcycling. Trained with LoRA in three stages — base SFT, a behavioural correction pass, and a reasoning-trace and tool-use distillation pass.

License

Apache 2.0.