Give it a goal and a toolbox — it plans the sequence, calls the right tools with the right arguments, reads every result, recovers from failures, and reports back in the user's own language.

Multi-step orchestration — chains 4+ tool calls per task without hand-holding
Strict, valid JSON — arguments match the schema every step
Multilingual by default — English, हिंदी, and Hinglish in, matching language out
Grounded answers — final confirmations quote real values from tool results (order IDs, totals, ETAs)
Graceful failure handling — surfaces errors and proposes alternatives instead of hallucinating success

Verified episode 1 — end-to-end order (English)

Unedited transcript, temperature 0.1. One request → four correct tool calls → grounded confirmation:

text

USER: Order 2 packets of Amul milk and 1 brown bread on Zepto. Pay cash on delivery.

[1] → search_multiple_products({"queries": ["Amul milk", "brown bread"]})
[2] → update_cart({"items": [{"product_id": "ZP-MILK-001", "quantity": 2},
                             {"product_id": "ZP-BREAD-007", "quantity": 1}]})
[3] → get_payment_methods({})
[4] → create_order({"payment_method": "COD"})

MODEL: Your order has been confirmed! 🛒
       Order ID: ZO-88121
       Items: 2 × Amul Taaza Toned Milk 500ml (₹29 each)
              1 × Britannia Brown Bread 400g (₹50)
       Total: ₹108 | Payment: Cash on Delivery
       Estimated delivery: 9 minutes.

Every value in the confirmation (order ID, prices, ETA) is read from tool results — nothing invented.

Verified episode 2 — same task, in Hinglish

text

USER: yaar zepto se 2 amul doodh aur ek brown bread mangwa de, cash on delivery rakhna

[1] → search_multiple_products({"queries": ["Amul doodh", "brown bread"]})
[2] → update_cart({"items": [{"product_id": "ZP-MILK-001", "quantity": 2},
                             {"product_id": "ZP-BREAD-007", "quantity": 1}]})
[3] → get_payment_methods({})
[4] → create_order({"payment_method": "COD"})

MODEL: Your order has been confirmed! 🛒  Order ID: ZO-88121 · Total ₹108 ·
       COD · delivery in ~9 minutes.

Note the search query: the model kept the user's own word — "Amul doodh" — and the catalog still resolved. Casual romanized Hindi is a first-class input, not an edge case.

Verified episode 3 — Hindi in, Hindi out

text

USER: कृपया Zepto से 2 अमूल दूध और 1 ब्राउन ब्रेड ऑर्डर करें। भुगतान कैश ऑन डिलीवरी होगा।

[1] → search_multiple_products({"queries": ["Amul milk", "brown bread"]})
[2] → update_cart({...})     [3] → get_payment_methods({})     [4] → create_order({"payment_method": "COD"})

MODEL: आपका ऑर्डर स्वीकृत हो गया है!
       ऑर्डर ID: ZO-88121
       कुल राशि: ₹108 · भुगतान विधि: कैश ऑन डिलीवरी (COD)
       अनुमानित डिलीवरी समय: 9 मिनट
       हम जल्द से जल्द आपके पास डिलीवर करेंगे! 🚚

Devanagari request → identical tool plan → full Devanagari confirmation. No language flag, no prompt engineering — the model mirrors the user.

Verified episode 4 — error recovery

The backend rejects COD mid-flow. The model does the right thing for a payments flow: it stops, explains, and asks — instead of silently switching or pretending success.

text

USER: Order 2 Amul milk on Zepto, cash on delivery.

[1] → search_multiple_products({"queries": ["Amul milk"]})
[2] → update_cart({"items": [{"product_id": "ZP-MILK-001", "quantity": 2}]})
[3] → create_order({"payment_method": "cod"})
      ← {"status": "error", "reason": "COD not available for this store. Available: UPI, Zepto Cash"}

MODEL: The Cash on Delivery (COD) option is not available for this store. You can
       proceed using UPI or Zepto Cash. Would you like to update your payment preference?

Measured results

All episodes above are unedited runs against the released weights (temperature 0.1), using a real quick-commerce tool set (search / cart / payments / ordering).

What we measured	Result
Multi-step episodes completed correctly	4 / 4 (English, Hinglish, Hindi, error-injection)
Tool selection across all steps	13 / 13 calls — correct tool every time
Argument validity (schema-conformant JSON)	13 / 13 calls
Language fidelity (reply matches user's language)	3 / 3
Hallucinated values in final answers	0 — every figure traceable to a tool result
Generation throughput (2× NVIDIA A40, released GGUF build)	~26.7 tok/s sustained across all 15 generations

Quick start

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-32B", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "jatayulabs/Qwen3-32B-ToolCall")
tok = AutoTokenizer.from_pretrained("jatayulabs/Qwen3-32B-ToolCall")

Prefer an OpenAI-compatible endpoint on a single GPU? Use the GGUF build — one llama-server command, full tools API support.

Checkpoints in this repo

Location	Step	Train loss	Notes
root / `checkpoint-600/`	600	0.070	✅ released — best checkpoint
`checkpoint-400/`	400	0.229	conservative early checkpoint

🔬 How we trained it

Trained on NVIDIA A100 GPUs with LLaMA-Factory, using LoRA on top of Qwen3-32B.

The curve tells the story: the model spends epoch 1 learning the format of structured tool calls (plateau ~0.45), masters it in a sharp drop at the epoch boundary (0.44 → 0.23), grinds through epoch 2 refining planning, then locks in at 0.070 by step 600. Training continued to step 800 but began to overfit (0.077) — step 600 is the release.

Setting	Value
Base model	Qwen3-32B (32.8B params)
Method	LoRA — rank 64, alpha 128, all attention + MLP projections
Trainable parameters	537M (1.6% of total)
Optimizer	LoRA+ (AdamW, lr-ratio 16)
Learning rate	1e-4, cosine decay, 10% warmup
Effective batch size	16 (2 × 8 grad accumulation)
Sequence length	2,048
Training data	5,000 curated multi-step tool-calling conversations — quick-commerce, payments, messaging, search — with English, Hindi, and Hinglish coverage
Final train loss	0.070 (step 600, epoch 2.1)

🪶 About Jatayu Labs

Jatayu Labs

Jatayu Labs is an Indian AI lab building agentic intelligence for a billion users — in their own languages.

🔧 Fine-tuned LLMs for agentic tool calling and orchestration
🎯 Specialised small language models for focused production tasks
🎙️ Voice-modality models — speech in, speech out
🇮🇳 Indic languages first — Hindi, Hinglish, and regional languages as first-class citizens

Jatayu — the bird of Indian epic who flew fastest and acted first.

License: Apache 2.0 · Contact: via the Community tab