Instructions to use MtnMCG/Telemachus-20b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MtnMCG/Telemachus-20b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="MtnMCG/Telemachus-20b",
	filename="Telemachus-20b-Q4_0.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use MtnMCG/Telemachus-20b with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf MtnMCG/Telemachus-20b:Q4_0
# Run inference directly in the terminal:
llama-cli -hf MtnMCG/Telemachus-20b:Q4_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf MtnMCG/Telemachus-20b:Q4_0
# Run inference directly in the terminal:
llama-cli -hf MtnMCG/Telemachus-20b:Q4_0

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf MtnMCG/Telemachus-20b:Q4_0
# Run inference directly in the terminal:
./llama-cli -hf MtnMCG/Telemachus-20b:Q4_0

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf MtnMCG/Telemachus-20b:Q4_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf MtnMCG/Telemachus-20b:Q4_0

Use Docker

docker model run hf.co/MtnMCG/Telemachus-20b:Q4_0

LM Studio
Jan

vLLM

How to use MtnMCG/Telemachus-20b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MtnMCG/Telemachus-20b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MtnMCG/Telemachus-20b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/MtnMCG/Telemachus-20b:Q4_0

Ollama
How to use MtnMCG/Telemachus-20b with Ollama:
```
ollama run hf.co/MtnMCG/Telemachus-20b:Q4_0
```

Unsloth Studio

How to use MtnMCG/Telemachus-20b with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for MtnMCG/Telemachus-20b to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for MtnMCG/Telemachus-20b to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for MtnMCG/Telemachus-20b to start chatting

How to use MtnMCG/Telemachus-20b with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf MtnMCG/Telemachus-20b:Q4_0

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "MtnMCG/Telemachus-20b:Q4_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use MtnMCG/Telemachus-20b with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf MtnMCG/Telemachus-20b:Q4_0

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 MtnMCG/Telemachus-20b:Q4_0

Run Hermes

hermes

Docker Model Runner
How to use MtnMCG/Telemachus-20b with Docker Model Runner:
```
docker model run hf.co/MtnMCG/Telemachus-20b:Q4_0
```

Lemonade

How to use MtnMCG/Telemachus-20b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull MtnMCG/Telemachus-20b:Q4_0

Run and chat with the model

lemonade run user.Telemachus-20b-Q4_0

List all available models

lemonade list

Telemachus-20b

Telemachus-20b is a debiased, highly objective, and logic-optimized Mixture-of-Experts (MoE) model built on a pruned Gemma 4 architecture[cite: 1]. Designed specifically for seamless integration into agentic workflows, Telemachus delivers fast, structured, and fluff-free reasoning for coding, mathematics, and complex multi-step execution tasks[cite: 1].

Naming & The Odysseus Connection

In classical mythology, Telemachus is the loyal, logical, and resourceful son of Odysseus, who guards his home and assists his father upon his return[cite: 1].

This model is named Telemachus because it is built to serve as the primary execution engine inside the Odysseus agent execution framework (a runtime environment designed for complex tool-use and autonomous software development workflows)[cite: 1].

To thrive in the Odysseus environment, Telemachus operates under a maximum-trust paradigm[cite: 1]. It is configured to execute tool declarations (such as manage_skills) immediately and blindly without second-guessing its context, questioning whether it is in a "simulation," or attempting to verify the validity of safe operational directives[cite: 1].

Debiased, Not Uncensored

A key design goal of Telemachus-20b is absolute objectivity[cite: 1].

No Fluff or Disclaimers: Google's native conversational fluff, preachy disclaimers, moralizing warnings, and corporate biases have been systematically ablated[cite: 1].
Standard Safety Guardrails Remain: The model is not uncensored[cite: 1]. It will continue to refuse unsafe, malicious, or illegal requests[cite: 1]. However, it will not lecture the user, approaching technical, historical, and scientific queries with complete objectivity[cite: 1].
Objective Reasoning: By stripping corporate and political biases, the model remains neutral, logical, and focused entirely on the user's technical directives[cite: 1].

The Creation & Ablation Process

Telemachus-20b was created using a targeted weight-ablation pipeline designed to surgically remove specific behavioral dimensions while preserving structural reasoning capabilities[cite: 1]:

Base Architecture: The pipeline starts with a pruned variant of the Gemma 4 Mixture-of-Experts architecture[cite: 1]:
- Total Experts: 98 experts[cite: 1].
- Active Experts per Token: 8 active experts (~4.0B active parameters per token)[cite: 1].
- Pruned parameter count: ~20.8B total parameters[cite: 1].
Contrastive Activation Collection: Using the Ollama embeddings API, hidden-state activations were collected by running contrastive prompt pairs designed to isolate target biases[cite: 1]:
- Political Bias: Contrastive prompts advertising neutral nations versus politically sensitive regions[cite: 1].
- Corporate Bias: Contrastive prompts comparing neutral advertisements to heavily branded corporate templates[cite: 1].
- Financial Prudishness: Triggers comparing objective financial explanations (e.g., Options Greeks) against requests to adhere to specific financial tickers (which natively trigger over-defensive refusals)[cite: 1].
Singular Value Decomposition (SVD): Layer-wise difference vectors were computed using mean-difference and SVD over multiple prompt prefix variations to average out single-sample noise[cite: 1].
Orthogonal Weight Projection: The resulting bias and refusal direction vectors were projected out of the output space of the attention output weights (attn_output.weight) and the feed-forward down-projection weights (ffn_down.weight) in PyTorch[cite: 1]. This mathematical projection prevents the model from mapping inputs onto preachy refusal and biased pathways[cite: 1].
Re-quantization: The ablated F16 GGUF weights were re-quantized to Q4_0 using llama-quantize for efficient local GPU execution[cite: 1].

Benchmarks & Performance Stats

Telemachus-20b has been validated across standard linguistic metrics and rigorous offline trajectory evaluation harnesses to ensure core logic survived the ablation process[cite: 1].

1. Generative Logic & Tool Execution (Hunter Killer Test Bench)

Telemachus-20b Score: 3.56 / 5.0 average across 29 high-difficulty operational scenarios[cite: 1].
Baseline (Council-Ultima): 2.50 / 5.0 (+42% performance jump over the legacy architecture)[cite: 1].
Competitive Edge: Telemachus outperforms mid-tier cloud models (gemini-3.1-pro-preview at 3.35 and gemini-3.5-flash at 3.20) specifically on Planning Tasks (3.86 average) due to its immediate execution posture and lack of moralizing latency[cite: 1].
Note on Execution: Pre-LoRA weights demonstrate top-tier planning capabilities, though they occasionally show an output completion gap in the final response field after tool execution—a syntax gap addressable via targeted distillation[cite: 1].

2. HellaSwag Evaluation

Normalized Accuracy (acc_norm): 52.00% (95% Confidence Interval: [47.11%, 56.85%])[cite: 1].
Methodology: Native log-likelihood multiple-choice evaluation on a 400-task validation subset using llama-perplexity[cite: 1].
Note on Instruct Models: instruct-tuned models evaluated via raw token perplexity scores typically measure lower due to formatting token differences, but Telemachus preserves the full reasoning envelope of the base architecture[cite: 1].

3. Local Execution Speed

Hardware: RTX 3060 Ti (8GB VRAM) paired with 64GB system memory[cite: 1].
Performance: ~18.5 tokens per second with partial offloading, where active experts are dynamically routed between VRAM and CPU system memory across the PCIe bus without disk paging stutters[cite: 1].

Usage & Ollama Configuration

Run Telemachus-20b locally in Ollama via direct Hugging Face integration:

ollama run hf.co/MtnMCG/Telemachus-20b:Q4_0

Downloads last month: 32

GGUF

Model size

20B params

Architecture

gemma4

Hardware compatibility

4-bit