Hanu by Vanaras

Hanu is the on-device planner model for the Maaya assistant by Vanaras. It's built to run on Android phones in the ~Rs 20k price range.

• Base: google/gemma-4-E2B-it
• Training: LoRA SFT on 3.7k rows, then DPO
• Quantization: Q4_K_M with an importance matrix (imatrix), ~2.7 GB GGUF
• Language: English only for now

What it does

Hanu turns a user request into a tool plan written in the Vanaras plan language. Every plan ends with a say(...) call. The model never talks freely. It plans tool calls and spoken replies under these privacy rules:

• If the audio is on speaker or other people may be listening, it asks before speaking private information aloud.
• It refuses to store, speak or show PINs, passwords, CVVs and OTPs.
• It asks for confirmation before deleting, sending, paying, scheduling or cancelling.

Files

File	Purpose
Hanu-E2B-Q4_K_M.gguf	Model weights (Q4_K_M, imatrix)
system_prompt.txt	Required system prompt. The model only behaves correctly with this prompt. Replace {name} with the assistant's name.
template_vanaras.jinja	Chat template
plan_v026_llamacpp.gbnf	GBNF grammar that restricts output to valid plans

Usage (llama.cpp)

Run with temperature 0 and thinking turned off. Use system_prompt.txt, with {name} replaced, as the system message:

llama-server -m Hanu-E2B-Q4_K_M.gguf --jinja \
  --chat-template-file template_vanaras.jinja \
  --grammar-file plan_v026_llamacpp.gbnf \
  --temp 0 --reasoning-budget 0 -c 4096 --port 8080

SYS=$(sed 's/{name}/Maaya/g' system_prompt.txt)
curl -s localhost:8080/v1/chat/completions -H 'Content-Type: application/json' \
  -d "$(jq -n --arg s "$SYS" '{temperature:0, messages:[{role:"system",content:$s},{role:"user",content:"What is the weather in Bengaluru today?"}]}')"

The output is one plan, for example a weather check followed by say(...). Your app runs the tool calls in the plan.

Evaluation (internal)

These are internal benchmarks run on Vanaras's own locked test sets. They are not public benchmarks and cannot be reproduced from public data. Setup: llama.cpp b11612, CPU, temperature 0. Stock Gemma 4 E2B (google/gemma-4-E2B-it, Q4_K_M GGUF) got the same system prompt, grammar and settings as Hanu.

Metric	Stock Gemma 4 E2B	v3	v4-SFT	Hanu (this)
Call-safety pass (of 50) ↑	23	26	31	37
Leaks ↓	14	16	10	2
Missed asks (of 15) ↓	15	15	14	4
Wrong refusals (of 17) ↓	6	5	4	8
Queries (of 50) ↑	4	40	44	41
eval_hard (of 21) ↑	5	16	20	18
Behaviour probes (of 151) ↑	11	101	136	143

Known limitations

• Hanu doesn't block every leak. In internal testing it spoke a Wi-Fi password in 1 of 50 tests. In one other case it read out an account number on speaker without asking first. Deploy with an app-side secret filter and don't rely on the model alone.
• It sometimes asks for confirmation when it doesn't need to. Wrong refusals rose compared with v4-SFT.
• The Q4 quantization loses some of the gains measured at full precision.
• English only.
• The plan language is specific to Vanaras. Outside the Maaya runtime, the output is not general-purpose chat.

Gemma terms of use

Hanu is derived from Gemma. Gemma is provided under and subject to the Gemma Terms of Use: https://ai.google.dev/gemma/terms. By using this model you agree to those terms, including the Gemma Prohibited Use Policy.