Instructions to use CinderellaMan/Flint-Pascal-Q8-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use CinderellaMan/Flint-Pascal-Q8-GGUF 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 CinderellaMan/Flint-Pascal-Q8-GGUF
# Run inference directly in the terminal:
llama cli -hf CinderellaMan/Flint-Pascal-Q8-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf CinderellaMan/Flint-Pascal-Q8-GGUF
# Run inference directly in the terminal:
llama cli -hf CinderellaMan/Flint-Pascal-Q8-GGUF

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 CinderellaMan/Flint-Pascal-Q8-GGUF
# Run inference directly in the terminal:
./build/bin/llama cli -hf CinderellaMan/Flint-Pascal-Q8-GGUF

Use Docker

docker model run hf.co/CinderellaMan/Flint-Pascal-Q8-GGUF

Llama new
LM Studio
Jan

vLLM

How to use CinderellaMan/Flint-Pascal-Q8-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "CinderellaMan/Flint-Pascal-Q8-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "CinderellaMan/Flint-Pascal-Q8-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/CinderellaMan/Flint-Pascal-Q8-GGUF

Ollama
How to use CinderellaMan/Flint-Pascal-Q8-GGUF with Ollama:
```
ollama run hf.co/CinderellaMan/Flint-Pascal-Q8-GGUF
```
Unsloth Desktop

How to use CinderellaMan/Flint-Pascal-Q8-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf CinderellaMan/Flint-Pascal-Q8-GGUF

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": "CinderellaMan/Flint-Pascal-Q8-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use CinderellaMan/Flint-Pascal-Q8-GGUF with Docker Model Runner:
```
docker model run hf.co/CinderellaMan/Flint-Pascal-Q8-GGUF
```

Lemonade

How to use CinderellaMan/Flint-Pascal-Q8-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull CinderellaMan/Flint-Pascal-Q8-GGUF

Run and chat with the model

lemonade run user.Flint-Pascal-Q8-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use CinderellaMan/Flint-Pascal-Q8-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf CinderellaMan/Flint-Pascal-Q8-GGUF

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 CinderellaMan/Flint-Pascal-Q8-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use CinderellaMan/Flint-Pascal-Q8-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf CinderellaMan/Flint-Pascal-Q8-GGUF

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 "CinderellaMan/Flint-Pascal-Q8-GGUF" \
  --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"

Flint Pascal Q8 GGUF: Flint at Q8_0's quality, 11 % smaller, for old NVIDIA cards and mining rigs

What this is. Flint is a 39.8-billion-parameter AI model that thinks before it answers, writes code and calls tools, and reads up to 262,144 tokens of context. This file (its internal name is p065) is Unsloth's Q8_0 of Flint with the tensors that cost the least quality moved to 5-bit Q5_0: 40.4 GB against Q8_0's 42.3 GB, and 11.0 % fewer bytes read per generated token. It is made for old NVIDIA "Pascal" cards (compute capability 6.1: the P106 and P104 cards of retired mining rigs, the GTX 10 series, the Tesla P4/P40); its speed is measured on P106-090 mining cards only.

Why it is faster there. A Pascal mining card spends a little over half of each generated token reading weights from its memory. Fewer bytes per token means less of that: in a paired replay of the same agent sessions this file generated each token in 0.941 of Q8_0's time, at the same prompt speed. Its KL divergence against the original 16-bit model is within two standard errors of Q8_0's.

How fast. A coding agent fixing twelve small bugs on a Manli mining rig with nine P106-090 cards (6 GB each, every card on a single PCIe 1.0 lane, a 2-core Celeron host with 4 GB of RAM), the full 262,144-token context: 47.4 tokens per second, all twelve tasks passed, on the Miner Llama v3 release build, against 45.9 for Unsloth's Q8_0 on the same build.

Which Flint file should you pick?

Nine 6 GB cards, best quality: this file or Unsloth's Q8_0 (42.3 GB). Both fit the full context on nine cards; this one is Q8_0's quality at fewer bytes per token.
Six to eight cards, or the most speed on nine: Flint Pascal Q5 GGUF (29.4 GB), at a lower quality.
On a modern GPU or a Mac: Unsloth's Q8_0. This file is measured on Pascal cards only.

Run it in three steps

Driver: Ubuntu 24.04 or newer with NVIDIA's driver 580, the last one for Pascal cards (setup guide).
Download the Pascal engine, Miner Llama v3.0, from the fork's releases, unpack it, and download flint-pascal-Q8.gguf from this page.

Start it, on nine cards with the full context (v2's release line, Flint's v3 preset):

llama-p106-v3.0/llama-server -m flint-pascal-Q8.gguf --alias flint -ngl 99 -fa on -c 262144 -np 1 -b 512 -ub 128 -ts 6,6,5,5,6,5,5,6,6 -bs --ctx-checkpoints 4 --cache-ram 0 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0 --pipe-depth 4 --host 127.0.0.1 --port 8080

Then open http://127.0.0.1:8080 through an SSH tunnel from your PC (ssh -N -L 8080:127.0.0.1:8080 <your rig>), or point any OpenAI-compatible client at it.

What each flag does and how to use it for coding agents: the Flint page; the v3 release: release/flint-v3.

What you need

About 41 GB of GPU memory plus room for the context: nine 6 GB cards for the full 262,144 tokens.
Miner Llama for the speed. The file uses only llama.cpp's own types (Q8_0, Q5_0, F16), so stock llama.cpp and apps built on it load it too, without the Pascal kernels.

Speed

The agent check gives the pi coding agent twelve small bug-fix tasks (Python and JavaScript, real test runs, the context growing every turn) and records every turn. Nine P106-090, 262,144-token context, the command line above, one run per row:

Engine and file	Tasks passed	Generation, tokens/s	Prompt reading, tokens/s	First token per turn (median)	Data
Stock llama.cpp b12076 (`20a51f24a`), Unsloth Q8_0, 65,536-token context, 2026-11-05 and 2026-11-06	10 of 12	14.4	148	4.13 s	2026-11-05_release_regression
Miner Llama v3 (`1f7215948`), Unsloth Q8_0, 2026-11-05	12 of 12	45.9	168	2.67 s	2026-11-05_release_regression
Miner Llama v3 (`1551f5fe1`), this file, 2026-11-06	12 of 12	47.4	175	2.66 s	2026-11-05_release_regression
Miner Llama v2.0 (`37e717ff6`), this file, 2026-10-28	12 of 12	47.1	—	—	2026-10-27_flint_near_q8

1f7215948 gives the same greedy output as the release build 1551f5fe1.
On the release build this file generates 3.3 % faster than Unsloth's Q8_0 (47.4 against 45.9 tokens/s), with a median task of 1.0 against 1.2 minutes. Its sessions wrote more text, so all twelve tasks took 22.3 against 19.2 minutes.
Stock runs at 65,536 tokens of context, because at 262,144 its host-side attention mask does not fit in 4 GB, on the line a llama-bench pass chose for it (-b 2048 -ub 512 -fa on, the same split). Its two misses ran into the 12-minute cap.
The paired replay (v2.0, the same six recorded agent sessions sent to Q8_0 and then to this file, replies capped at their recorded lengths): 0.941 of Q8_0's time per generated token (95 % interval 0.930–0.980 over the sessions), 1.002 per prompt token, 0.963 of the time for the same work.

Quality

Against Flint's original 16-bit weights, 80 × 512 tokens of wikitext-2, all three files on the same GPU set:

File	Size	Bytes read per token	Picks the same next word as the original	Distance from the original (KL divergence, lower is better)
Unsloth Q8_0	42.3 GB	2.696 GB	92.39 ± 0.19 %	0.0627 ± 0.0032
This file	40.4 GB	2.400 GB	91.78 ± 0.19 %	0.0656 ± 0.0029
Flint Pascal Q5 GGUF	29.4 GB	2.090 GB	88.44 ± 0.22 %	0.1108 ± 0.0037

On coding-agent text (12 × 8,192 tokens, against Q8_0) this file scores 0.2292 and 93.42 %; Q8_0 with only its routers in F16, the floor for every file with F16 routers, scores 0.2273 and 93.78 %.
The engine keeps the answers: its kernels change only the order in which numbers are added up, and pipelined speculation gives exactly the tokens of plain decoding.

How I made it, and what changed

Every tensor of Flint was given a cost: the KL divergence that moving it alone from Q8_0 to Flint Pascal Q5 GGUF's Q5_0 adds, by tensor class and layer group. Dynamic programming over the bytes each move saves per token picked the set with the least added divergence at each size. Of the three picks composed and scored, this file saves the most while its KL divergence stays within noise of Q8_0's. The changes against Unsloth's Q8_0 (revision 20468edf), in the GGUF format:

The routers in 16-bit floats (F16), taken from Flint Pascal Q5 GGUF.
In Q5_0, taken from Flint Pascal Q5 GGUF: attention q, k, v and output in layers 33–48, the shared expert in layers 33–48, the routed experts' gate and up in layers 41–48, and the output head (Q5_0 rounded with GPTQ).
Everything else is Unsloth's Q8_0, unchanged. Early layers stay at Q8_0: they carry most of the cost.

The scores of every move, the picks and the composing script: 2026-10-27_flint_near_q8.

Licence and credits

Flint is made by RecursiveSI and released under Apache-2.0; this file is a quantized derivative under the same licence (LICENSE), with the changes listed above.
Built from Unsloth's Flint Q8_0 (Apache-2.0); Unsloth's importance matrix guided the Q5_0 tensors' rounding.
Quantized by Kamil Skowron (@kamilskowron on X), for LLMs on crypto-mining rigs. Not affiliated with RecursiveSI or Unsloth.

File	Size	SHA-256
`flint-pascal-Q8.gguf`	40.4 GB	see CHECKSUMS

Check a download with sha256sum -c CHECKSUMS.

Downloads last month: 269

GGUF

Model size

40B params

Architecture

flint

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for CinderellaMan/Flint-Pascal-Q8-GGUF

Base model

RecursiveSI/flint

Quantized

(102)

this model