GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Novasaki
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
Novasaki
Model Tree
HuggingFaceTB/SmolLM2-1.7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
| Benchmark Domain | Native Normal Baseline | Grugified Adapter | Absolute Gain | Relative Surge |
|---|---|---|---|---|
| Tool Calling (Glaive AI 100) | 5.0% | 95.0% | +90.0% | 19.0x Surge 🚀 |
| Quantitative Math (GSM8K 100) | 0.0% | 9.0% | +9.0% | Non-zero logic |
| Coding (HumanEval 100) | 8.0% | 11.0% | +3.0% | 1.38x |
| Overall 300-Item Benchmark | 4.3% | 38.3% | +34.0% | 8.9x Surge |
Compact models (< 2B) lack the working memory to maintain both conversational self-debate and strict JSON schemas. Under native unconstrained prompting, SmolLM2 outputs rambling natural language, exhausting context before emitting valid JSON.
Grug Speech pins down the exact target invariants:
text
<think>Domain: tool_useGoal: get weatherInvariants: city="Tokyo", unit="celsius"Action: emit get_current_weatherDone.</think><tool_call>{"name": "get_current_weather", "arguments": {"city": "Tokyo", "unit": "celsius"}}</tool_call>Result: 95 out of 100 tool tasks passed successfully.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel BASE_MODEL = "HuggingFaceTB/SmolLM2-1.7B-Instruct"ADAPTER_ID = "Novasaki/SmolLM2-1.7B-UniversalGrugifier" tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)base_model = AutoModelForCausalLM.from_pretrained( BASE_MODEL, torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base_model, ADAPTER_ID)FULL_TECHNICAL_REPORT_GRUG_SPEECH_REASONING.txt<think>Domain: tool_useGoal: get weatherInvariants: city="Tokyo", unit="celsius"Action: emit get_current_weatherDone.</think><tool_call>{"name": "get_current_weather", "arguments": {"city": "Tokyo", "unit": "celsius"}}</tool_call>import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel BASE_MODEL = "HuggingFaceTB/SmolLM2-1.7B-Instruct"ADAPTER_ID = "Novasaki/SmolLM2-1.7B-UniversalGrugifier" tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)base_model = AutoModelForCausalLM.from_pretrained( BASE_MODEL, torch_dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base_model, ADAPTER_ID)