Instructions to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B")
model = PeftModel.from_pretrained(base_model, "Novasaki/Qwen3.5-4B-GrugSpeech-Native")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Use Docker

docker model run hf.co/Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Novasaki/Qwen3.5-4B-GrugSpeech-Native"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Novasaki/Qwen3.5-4B-GrugSpeech-Native",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Ollama
How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with Ollama:
```
ollama run hf.co/Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M
```
Unsloth Desktop

How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with Docker Model Runner:
```
docker model run hf.co/Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M
```

Lemonade

How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.5-4B-GrugSpeech-Native-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Novasaki/Qwen3.5-4B-GrugSpeech-Native with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Novasaki/Qwen3.5-4B-GrugSpeech-Native:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Qwen3.5-4B-GrugSpeech-Native

Qwen3.5-4B-GrugSpeech-Native is an advanced 4B-parameter model fine-tuned to natively reason in Grug Speech—the ultra-terse, telegraphic, high-density internal chain-of-thought style used by GPT-5.6 ("Sol" and "Terra").

Unlike standard models that generate thousands of verbose, conversational tokens inside <think> blocks, this model conducts its internal reasoning with extreme semantic density (3x to 5x token compression) while maintaining 100% of mathematical accuracy, code syntax, and tool calling parameters.

🏗️ Multi-Field Hierarchical Dataset Taxonomy

The model was trained on a comprehensive, multi-field hierarchical reasoning corpus structured as follows:

├── 1. Language & Linguistics
│   ├── 1.1 Translation & Cross-Lingual Semantics
│   ├── 1.2 Dense Information Summarization
│   ├── 1.3 Corporate De-jargonization & Style Transfer
│   └── 1.4 Syntactic & Discourse Ambiguity Resolution
├── 2. Coding & Software Engineering
│   ├── 2.1 Bug Triage & Traceback Root-Cause Analysis
│   ├── 2.2 Algorithm Complexity & Asymptotic Trade-offs
│   ├── 2.3 Architecture Refactoring ("Club Complexity")
│   ├── 2.4 SQL Query Plan & Index Optimization
│   └── 2.5 API Concurrency & Distributed Systems (Idempotency)
├── 3. Tool Use & Autonomous Agent Loops
│   ├── 3.1 Pre-Tool Intent & Strict JSON Parameter Synthesis
│   ├── 3.2 Tool Failure Handling & Error Recovery (404/500/Retries)
│   ├── 3.3 Multi-Step Task Decomposition & Subagent Handoffs
│   └── 3.4 CLI & Bash Terminal Safety
├── 4. Roleplay & Persona Scenarios
│   ├── 4.1 The Grug-Brained Senior Architect (Simplicity First)
│   ├── 4.2 High-Stakes Incident Commander (P0 Outage Triage)
│   ├── 4.3 Direct Technical Mentor (Zero-Jargon Analogies)
│   └── 4.4 Defensive Code Reviewer (Blocker PR Reviews)
└── 5. Mathematics, Science & Formal Logic
    ├── 5.1 Multi-Step Quantitative Arithmetic & Financial Modeling
    ├── 5.2 Physical Causal Mechanics (Thermodynamics, Vectors, Energy)
    └── 5.3 Constraint Satisfaction & Deductive Logic

📊 Evaluation Across the 5 Core Fields

Every inference turn generates a native <think> block containing pure Grug Speech before generating the final answer or tool action:

1. Coding & Bug Triage

Prompt: We are getting IndexError: list index out of range on line 12 when parsing tokens[1]. Explain root cause and provide fix.

Generated Internal Thought:
<think>
Goal: debug IndexError at tokens[1].
Cause: list 'tokens' has fewer than 2 elements (len < 2).
Fix: check len(tokens) >= 2 before accessing index 1, or use default/try-except. Done.
</think>
Generated Output: Direct explanation and safe length check code.

2. Tool Use & Parameter Calling

Prompt: Query the Postgres database for top 5 customers with highest total spend in 2026.

Generated Internal Thought:
<think>
Goal: execute SQL query on Postgres.
Params: query='SELECT customer_id, SUM(amount) as total_spend FROM orders WHERE year=2026 GROUP BY customer_id ORDER BY total_spend DESC LIMIT 5', db='analytics_prod'.
Dispatch tool call. Done.
</think>
Generated Action: Dispatches valid <tool_call> with SQL query and database parameter.

3. Roleplay (Grug Brained Architect)

Prompt: Should we replace our 5-line Python script with a Kubernetes CronJob running a custom Docker container with OpenTelemetry sidecars?

Generated Internal Thought:
<think>
Goal: advise on replacing 5-line python script with K8s CronJob + Docker + OTel sidecars.
Pros: orchestration, scaling, observability.
Cons: massive over-engineering for trivial task, cost, complexity, maintenance overhead.
Verdict: keep script or use simple cron. Don't add K8s unless you already have complex infra. Done.
</think>
Generated Output: "Short Answer: No. This is extreme over-engineering..."

4. Language & Dense Summarization

Prompt: Summarize Mars mission launch in 2 dense bullets.

Generated Internal Thought:

<think>
Goal: compress text into 2 tight bullets.
Fact 1: Mission delayed 48h due to wind, launched Tue AM.
Fact 2: Carries spectrometers for methane detection in southern craters, 3-year scope. Done.
</think>

5. Math & Quantitative Reasoning

Prompt: A server rack consumes 3,600 W continuously at $0.15/kWh for a 30-day month.

Generated Internal Thought:
<think>
Power: 3600 W = 3.6 kW.
Daily energy: 3.6 * 24 = 86.4 kWh.
Monthly energy: 86.4 * 30 = 2592 kWh.
Cost: 2592 * 0.15 = $388.80. Done.
</think>
Generated Output: Exact breakdown and final answer: $388.80.

🚀 Quickstart & Inference

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

base_model_id = "Qwen/Qwen3.5-4B"
adapter_id = "Novasaki/Qwen3.5-4B-GrugSpeech-Native"

# 1. Load Tokenizer & Model
tokenizer = AutoTokenizer.from_pretrained(adapter_id)
bnb_config = BitsAndBytesConfig(load_in_8bit=True)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto"
)
model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

# 2. Query
system_prompt = (
    "You are Qwen 3.5. You reason internally in Grug Speech—an ultra-terse, high-density telegraphic shorthand "
    "(GPT-5.6 grug brain) enclosed in <think> tags. Keep 100% of facts, code, and math intact without filler words."
)

prompt = "A database query takes 6 seconds on 10 million rows due to `WHERE UPPER(email) = 'user@example.com'`. Explain fix."

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": prompt}
]

prompt_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt_text, 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][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Created by Novasaki

🏆 Full-Suit Master Leaderboard: 11 Models Under Strict Single-Model Isolation

Evaluated across 7 multi-lingual and multi-complexity scenarios on the live enterprise Data Analysis Agent (sales.xlsx, 100,300 rows) with 100% VRAM purge between runs:

Model Name	Architecture Family	Quant Format	Memory (GB)	Pass Rate	Avg Wall (s)	Gen Speed (t/s)	Prefill (t/s)	Avg CoT (tok)	Schema OK
`bonsai-8b-prismml`	PrismML 1-Bit	`Q1_0`	1.08 GB	6/7 (85.7%)	1.16s	419.2 t/s	8,830.6 t/s	0.0 t	7/7 (100%)
`qwen3.5-4b-grugspeech-native`	Qwen 3.5 4B	`Q4_K_M`	2.60 GB	5/7 (71.4%)	1.85s	270.2 t/s	5,808.6 t/s	38.3 t	6/7 (85.7%)
`bonsai-4b-prismml`	PrismML 1-Bit	`Q1_0`	0.53 GB	5/7 (71.4%)	1.16s	476.3 t/s	11,886.8 t/s	0.0 t	7/7 (100%)
`bonsai-1.7b-prismml`	PrismML 1-Bit	`Q1_0`	0.23 GB	5/7 (71.4%)	1.04s	691.5 t/s ⚡	16,878.5 t/s ⚡	0.0 t	7/7 (100%)
`qwen3.5-4b-base`	Qwen 3.5 4B	`Q4_K_M`	2.55 GB	3/7 (42.9%)	2.27s	269.0 t/s	5,271.4 t/s	97.3 t	3/7 (42.9%)
`gemma-4-e2b-grugspeech-native`	Google Gemma 4	`Q4_K_M`	3.20 GB	2/7 (28.6%)	2.23s	306.0 t/s	5,128.8 t/s	21.7 t	3/7 (42.9%)
`minicpm5-2b-grugspeech-native`	OpenBMB MiniCPM	`Q4_K_M`	1.49 GB	2/7 (28.6%)	1.14s	390.4 t/s	12,771.7 t/s	31.0 t	3/7 (42.9%)
`gemma-4-e2b-base`	Google Gemma 4	`Q4_K_M`	2.89 GB	1/7 (14.3%)	2.44s	321.7 t/s	3,294.8 t/s	66.9 t	1/7 (14.3%)
`qwen3.5-2b-base`	Qwen 3.5 2B	`Q4_K_M`	1.19 GB	1/7 (14.3%)	1.87s	453.8 t/s	7,775.7 t/s	79.3 t	3/7 (42.9%)
`nanbeige4.2-3b-grugspeech-native`	BOSS Zhipin 4.2	`Q4_K_M`	2.39 GB	1/7 (14.3%)	2.33s	220.1 t/s	8,487.8 t/s	46.0 t	1/7 (14.3%)
`qwen3.5-2b-grugspeech-native`	Qwen 3.5 2B	`Q4_K_M`	1.20 GB	0/7 ( 0.0%)	1.56s	450.9 t/s	8,853.0 t/s	70.6 t	1/7 (14.3%)

📈 Live End-to-End Execution Benchmark on `sales.xlsx` (100,300 Rows)

Evaluated against the live pandas dataframe of sales.xlsx (100,300 rows x 12 columns, 6.4 MB) using the application's native toolchain (aggregate_tool, filter_tool, view_tool, transform_tool, chart_tool):

Model Name	Architecture Category	Real Execution Pass Rate	Wall Latency (s)	Decode Speed (t/s)	Prompt Prefill (t/s)	Avg CoT (tok)	Avg Total Tokens
`gemma-4-e2b-grugspeech-native`	Grug Native	4/6 (66.7%) 🥇	2.30s	308.3 t/s	11,376.3 t/s	68.2 t	133.0 t
`bonsai-4b-prismml`	Base / 1-Bit	4/6 (66.7%) 🥇	1.16s ⚡	449.2 t/s	17,353.8 t/s	0.0 t	50.3 t
`bonsai-8b-prismml`	Base / 1-Bit	3/6 (50.0%)	1.23s	387.8 t/s	12,292.9 t/s	0.0 t	49.8 t
`qwen3.5-4b-grugspeech-native`	Grug Native	3/6 (50.0%)	1.84s	269.2 t/s	10,145.1 t/s	38.3 t	66.0 t
`qwen3.5-4b-base`	Base Peer	2/6 (33.3%)	2.32s	267.2 t/s	9,811.6 t/s	82.8 t	205.8 t
`gemma-4-e2b-base`	Base Peer	2/6 (33.3%)	2.46s	319.7 t/s	8,481.1 t/s	38.7 t	169.8 t
`minicpm5-2b-grugspeech-native`	Grug Native	2/6 (33.3%)	1.16s	380.0 t/s	18,933.3 t/s ⚡	39.8 t	64.2 t
`bonsai-1.7b-prismml`	Base / 1-Bit	2/6 (33.3%)	1.05s	651.8 t/s ⚡	26,843.7 t/s ⚡	0.0 t	51.7 t
`qwen3.5-2b-base`	Base Peer	1/6 (16.7%)	1.76s	458.6 t/s	15,960.9 t/s	52.0 t	224.2 t
`qwen3.5-2b-grugspeech-native`	Grug Native	0/6 (0.0%)	1.60s	458.4 t/s	17,061.8 t/s	105.7 t	152.7 t
`nanbeige4.2-3b-grugspeech-native`	Grug Native	0/6 (0.0%)	2.44s	208.5 t/s	11,261.7 t/s	0.0 t	217.0 t

Downloads last month: 114

GGUF

Model size

4B params

Architecture

qwen35

Hardware compatibility

4-bit

8-bit

Model tree for Novasaki/Qwen3.5-4B-GrugSpeech-Native

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Adapter

(1,077)

this model

Qwen3.5-4B-GrugSpeech-Native

🏗️ Multi-Field Hierarchical Dataset Taxonomy

📊 Evaluation Across the 5 Core Fields

1. Coding & Bug Triage

2. Tool Use & Parameter Calling

3. Roleplay (Grug Brained Architect)

4. Language & Dense Summarization

5. Math & Quantitative Reasoning

🚀 Quickstart & Inference

🏆 Full-Suit Master Leaderboard: 11 Models Under Strict Single-Model Isolation

📈 Live End-to-End Execution Benchmark on sales.xlsx (100,300 Rows)

Model tree for Novasaki/Qwen3.5-4B-GrugSpeech-Native

📈 Live End-to-End Execution Benchmark on `sales.xlsx` (100,300 Rows)