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.
Model Details
Model Provider
Novasaki
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
Evaluated under strict head-to-head empirical testing against base unconstrained reasoning:
| Benchmark Domain | Native Normal Baseline | Grugified Adapter | Absolute Gain | Latency / Speedup |
|---|---|---|---|---|
| Tool Calling (Glaive AI 100) | 44.0% | 78.0% | +34.0% | 1.35s (137 tokens) |
| Quantitative Math (GSM8K 100) | 3.0% | 24.0% | +21.0% (8x rel) | 13.0x speedup (156s o 12s) |
| Coding (HumanEval 100) | 14.0% | Cognitive Plan | Compiler mode | Zero schema drift |
| Overall 300-Item Benchmark | 20.3% | 34.0% | +13.7% | 47% token reduction |
sales.xlsx, 100,300 rows)filter_tool, aggregate_tool, view_tool, chart_tool).Instead of hundreds of tokens of discursive self-doubt, Grug Speech enforces strict invariant conservation:
text
<think>Domain: tool_useGoal: dispatch payment verificationInvariants: order_id=1001, amount=250.0, currency="USD"Action: call verify_payment(order_id=1001, amount=250.0, currency="USD")Done.</think><tool_call>{"name": "verify_payment", "arguments": {"order_id": 1001, "amount": 250.0, "currency": "USD"}}</tool_call>python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel BASE_MODEL = "Qwen/Qwen3.5-4B"ADAPTER_ID = "Novasaki/Qwen3.5-4B-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) prompt = """You are a data analysis planner. Reason in Grug Speech inside <think> tags.User Query: Filter orders where Total > 500, group by City, and sum Total.""" inputs = tokenizer(prompt, return_tensors="pt").to("cuda")with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=250, do_sample=False) print(tokenizer.decode(out[0], skip_special_tokens=True))For full theoretical derivations (EBFT energy function, length penalty γ=0.005, margin m=1.5), complete 1,800-task master scorecard, and reproduction instructions, see:
FULL_TECHNICAL_REPORT_GRUG_SPEECH_REASONING.txt<think>Domain: tool_useGoal: dispatch payment verificationInvariants: order_id=1001, amount=250.0, currency="USD"Action: call verify_payment(order_id=1001, amount=250.0, currency="USD")Done.</think><tool_call>{"name": "verify_payment", "arguments": {"order_id": 1001, "amount": 250.0, "currency": "USD"}}</tool_call>import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel BASE_MODEL = "Qwen/Qwen3.5-4B"ADAPTER_ID = "Novasaki/Qwen3.5-4B-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) prompt = """You are a data analysis planner. Reason in Grug Speech inside <think> tags.User Query: Filter orders where Total > 500, group by City, and sum Total.""" inputs = tokenizer(prompt, return_tensors="pt").to("cuda")with torch.no_grad(): out = model.generate(**inputs, max_new_tokens=250, do_sample=False) print(tokenizer.decode(out[0], skip_special_tokens=True))