Instructions to use GenerallyTso/Sophro-4B-memory-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use GenerallyTso/Sophro-4B-memory-v1 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 GenerallyTso/Sophro-4B-memory-v1:Q4_0
# Run inference directly in the terminal:
llama cli -hf GenerallyTso/Sophro-4B-memory-v1:Q4_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf GenerallyTso/Sophro-4B-memory-v1:Q4_0
# Run inference directly in the terminal:
llama cli -hf GenerallyTso/Sophro-4B-memory-v1:Q4_0

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 GenerallyTso/Sophro-4B-memory-v1:Q4_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf GenerallyTso/Sophro-4B-memory-v1:Q4_0

Use Docker

docker model run hf.co/GenerallyTso/Sophro-4B-memory-v1:Q4_0

Llama new
LM Studio
Jan

vLLM

How to use GenerallyTso/Sophro-4B-memory-v1 with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/GenerallyTso/Sophro-4B-memory-v1:Q4_0

Ollama
How to use GenerallyTso/Sophro-4B-memory-v1 with Ollama:
```
ollama run hf.co/GenerallyTso/Sophro-4B-memory-v1:Q4_0
```
Unsloth Desktop

How to use GenerallyTso/Sophro-4B-memory-v1 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf GenerallyTso/Sophro-4B-memory-v1:Q4_0

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": "GenerallyTso/Sophro-4B-memory-v1:Q4_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use GenerallyTso/Sophro-4B-memory-v1 with Docker Model Runner:
```
docker model run hf.co/GenerallyTso/Sophro-4B-memory-v1:Q4_0
```

Lemonade

How to use GenerallyTso/Sophro-4B-memory-v1 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull GenerallyTso/Sophro-4B-memory-v1:Q4_0

Run and chat with the model

lemonade run user.Sophro-4B-memory-v1-Q4_0

List all available models

lemonade list

Hermes Agent

How to use GenerallyTso/Sophro-4B-memory-v1 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf GenerallyTso/Sophro-4B-memory-v1: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 GenerallyTso/Sophro-4B-memory-v1:Q4_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use GenerallyTso/Sophro-4B-memory-v1 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf GenerallyTso/Sophro-4B-memory-v1:Q4_0

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 "GenerallyTso/Sophro-4B-memory-v1:Q4_0" \
  --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"

Sophro-4B-memory-v1

Sophro is the on-device model of Zynkbot, a personal AI assistant that keeps its memory of you on your own devices. It is Qwen3-4B-Instruct-2507 fine-tuned for one habit: when it reads what you said, it stores what was said and nothing more. The name is from sophrosyne, the Greek word for soundness of mind and self-restraint.

Zynkbot downloads this file at first run and uses it for chat, memory extraction and memory linking on phones and computers, with no key and nothing sent anywhere.

File

File	Format	Size	SHA-256
`Sophro-4B-memory-v1-Q4_0.gguf`	GGUF, Q4_0	2.37 GB	`da0a9b08e9be4f2a9d806b85f70ce90c824e8e0fb8fe53c79958a84e0c5f5b1b`

Q4_0 was chosen so that the same file serves the CPU, GPU and NPU paths on Snapdragon phones. Runs with llama.cpp and anything built on it. Context window as the base model; Zynkbot uses 4096 tokens on phones and 8192 on computers.

How it was trained

A LoRA adapter was trained on a single consumer GPU (about ten minutes a run) from two sources: 933 real user messages labelled by a stronger model (Claude) with the exact memory-extraction prompt Zynkbot sends, and 349 memory-relationship cases written by the same model, all filtered by the project's benchmark pass rules. The adapter was merged into the base model and the result quantised to Q4_0. The loss was taken on the answer only.

The training messages are one person's private conversations and are not published.

Results

Zynkbot's memory bench, 128 held-out cases that were never in the training data, phone prompt, 2026-10-28:

Model	Facts found (of 69)	Invented	Junk (of 26)	Relationships (of 44)	Extraction time	Relationship time
Base Qwen3-4B Q4_0	66	7	0	28	0.30 s	0.99 s
Sophro-4B-memory-v1	68	1	0	38	0.31 s	0.21 s

The six relationship misses are five "updates" called "contradicts" and one "none" called "related". Full method and the other benches: BENCHMARKS.md.

Standard benchmarks

Did the tuning cost the model anything else? Both models the exact Q4_0 files above, llama.cpp on one GPU, temperature 0, lm-evaluation-harness 0.4.13, 2026-11-02/03:

Benchmark	Base Qwen3-4B Q4_0	Sophro-4B-memory-v1
IFEval, prompt-level strict	81.3%	81.2%
IFEval, prompt-level loose	84.8%	85.4%
GSM8K, 8-shot CoT, strict match	80.0%	83.8%
MMLU, 25 per subject, scored by the first answer letter	71.6%	71.4%

Instruction following and knowledge are unchanged, arithmetic slightly better. Two honest notes: the harness's stock MMLU scorer wants the bare letter, and Sophro tends to write "D. 0,4" rather than "D", which the stock scorer counts as wrong (that line reads 54.7% against 24.0%; the first-letter score above is the one that measures knowledge). And on IFEval's 31 prompts that ask for a reply in another language, Sophro passes 23 against the base's 30: it tends to fall into repetition in Hindi, Marathi, Nepali and Bengali. On the 510 English prompts it is at 81.6% against 80.4%. Zynkbot uses it in English.

Prompt format

The Qwen3 chat template, as declared in the GGUF file. No thinking mode.

Licence

Downloads last month: 58

GGUF

Model size

4B params

Architecture

qwen3

Hardware compatibility

4-bit

Model tree for GenerallyTso/Sophro-4B-memory-v1

Base model

Qwen/Qwen3-4B-Instruct-2507

Quantized

(363)

this model