GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Horizon-Labs-Inc
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Horizon-Labs-Inc
Model Tree
Qwen/Qwen3-1.7B-Base
Input Modalities
Output Modalities
Supported Functionality
<tool_call>{"name": ..., "arguments": ...}</tool_call>),
so it works out of the box with the standard Qwen3 tool parsers in
transformers and vLLM (no custom parser needed).enable_thinking=False): deterministic, low-latency
tool calls. generation_config.eos_token_id includes both
and so generation stops cleanly at turn end.Exact-match evaluation with the released harness (scripts/eval_toolcall.py).
Full exact match = same set of function names and identical arguments.
| Benchmark | valid JSON | tool-name match | full exact match |
|---|---|---|---|
| xLAM holdout (1,000 rows never trained on) | 100% | 96.5% | 79.5% |
| ToolACE (external dataset, not trained on) | 100% | 95.5% | 75.3% |
Both sources are disjoint from training: the first 1,000 xLAM rows were held out, and ToolACE was never used at all. These are the authors' own harness numbers, not BFCL. Berkeley Function Calling Leaderboard (BFCL v4) evaluation is in progress and will be added here.
python
from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Horizon-Labs/qwen3-1.7b-toolsmith"tok = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto") tools = [{ "type": "function", "function": { "name": "get_weather", "description": "Get the current weather for a city", "parameters": { "type": "object", "properties": {"city": {"type": "string", "description": "City name"}}, "required": ["city"], }, },}]messages = [{"role": "user", "content": "What's the weather in Berlin?"}] inputs = tok.apply_chat_template( messages, tools=tools, add_generation_prompt=True, enable_thinking=False, return_tensors="pt",).to(model.device)out = model.generate(inputs, max_new_tokens=256, do_sample=False)print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))Expected output:
text
<tool_call>{"name": "get_weather", "arguments": {"city": "Berlin"}}</tool_call>Qwen/Qwen3-1.7B-Base-100), including the end-of-turn token.minpeter/xlam-function-calling-60k-parsed (CC-BY-4.0; derived from
Salesforce/xLAM, generated by APIGen) - 58,647: single, multiple and parallel
calls. First 1,000 rows reserved for evaluation.NousResearch/hermes-function-calling-v1 (Apache-2.0) - function calling
(1,100), JSON mode agentic (1,008), JSON mode single-turn (1,241):
structured extraction and multi-step tool use.HuggingFaceTB/smoltalk (Apache-2.0) - 7,950 general-instruction replay
(~11%) to limit capability regression.scripts/prepare_data.py, scripts/train_sft.py,
scripts/eval_toolcall.py and the exact command in train.sh.Intended for agentic tool use, structured extraction, and JSON generation where a small, locally-runnable model is preferred. It is not a general-purpose chat model and does not do open-ended reasoning.
Known limitations:
tools schema, keep catalogs
small, and verify before acting.Training data includes work by Salesforce (xLAM / APIGen, CC-BY-4.0), Nous Research (Hermes, Apache-2.0) and Hugging Face (smoltalk, Apache-2.0). Base model by the Qwen team. See the dataset cards for full lineage and terms.
If you use this model, please cite the base model and the datasets above.
from transformers import AutoModelForCausalLM, AutoTokenizer model_id = "Horizon-Labs/qwen3-1.7b-toolsmith"tok = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained(model_id, dtype="bfloat16", device_map="auto") tools = [{ "type": "function", "function": { "name": "get_weather", "description": "Get the current weather for a city", "parameters": { "type": "object", "properties": {"city": {"type": "string", "description": "City name"}}, "required": ["city"], }, },}]messages = [{"role": "user", "content": "What's the weather in Berlin?"}] inputs = tok.apply_chat_template( messages, tools=tools, add_generation_prompt=True, enable_thinking=False, return_tensors="pt",).to(model.device)out = model.generate(inputs, max_new_tokens=256, do_sample=False)print(tok.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))