Instructions to use amkkk/AMKemma-3.1-Write with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use amkkk/AMKemma-3.1-Write 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 amkkk/AMKemma-3.1-Write:Q8_0
# Run inference directly in the terminal:
llama cli -hf amkkk/AMKemma-3.1-Write:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf amkkk/AMKemma-3.1-Write:Q8_0
# Run inference directly in the terminal:
llama cli -hf amkkk/AMKemma-3.1-Write:Q8_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 amkkk/AMKemma-3.1-Write:Q8_0
# Run inference directly in the terminal:
./build/bin/llama cli -hf amkkk/AMKemma-3.1-Write:Q8_0

Use Docker

docker model run hf.co/amkkk/AMKemma-3.1-Write:Q8_0

Llama new
LM Studio
Jan

vLLM

How to use amkkk/AMKemma-3.1-Write with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "amkkk/AMKemma-3.1-Write"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "amkkk/AMKemma-3.1-Write",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/amkkk/AMKemma-3.1-Write:Q8_0

Ollama
How to use amkkk/AMKemma-3.1-Write with Ollama:
```
ollama run hf.co/amkkk/AMKemma-3.1-Write:Q8_0
```
Unsloth Desktop

How to use amkkk/AMKemma-3.1-Write with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf amkkk/AMKemma-3.1-Write:Q8_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": "amkkk/AMKemma-3.1-Write:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use amkkk/AMKemma-3.1-Write with Docker Model Runner:
```
docker model run hf.co/amkkk/AMKemma-3.1-Write:Q8_0
```

Lemonade

How to use amkkk/AMKemma-3.1-Write with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull amkkk/AMKemma-3.1-Write:Q8_0

Run and chat with the model

lemonade run user.AMKemma-3.1-Write-Q8_0

List all available models

lemonade list

Hermes Agent

How to use amkkk/AMKemma-3.1-Write with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf amkkk/AMKemma-3.1-Write:Q8_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 amkkk/AMKemma-3.1-Write:Q8_0

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use amkkk/AMKemma-3.1-Write with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf amkkk/AMKemma-3.1-Write:Q8_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 "amkkk/AMKemma-3.1-Write:Q8_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"

AMKemma - 3.1 Write

Abstract

AMKemma - 3.1 Write is a writing fine-tune of Gemma 4 E2B IT, released as a standalone Q8_0 GGUF. Give it a brief, a voice and a target length, and it writes from those instructions.

Training

Trained on brief-to-human-page pairs from AMKwen-3.0-Write-Corpus, using human-response cross-entropy + token-level unlikelihood, weighted equally. Cross-entropy learns the human continuation; unlikelihood penalizes earlier context tokens when they compete with the correct next token. Prompts are masked from the training targets, and native end-of-turn tokens are preserved.

rsLoRA: rank 16, alpha 32, dropout 0.05, learning rate 2e-5, 4096-token windows with 256-token overlap, gradient accumulation 4, across 277 text linear layers. This release uses checkpoint 42,831, selected by held-out cross-entropy, with the adapter merged before Q8 conversion.

Files and usage

Download AMKemma-3.1-Write-Q8_0.gguf (5.38 GB) and load it in a llama.cpp-compatible runtime with Gemma 4 support. Use the embedded Gemma chat template and native stops <turn|> / <eos> (106 / 1), with thinking off.

Starting settings: temperature 0.8, top-p 1, top-k 0, min-p 0, repetition penalty 1. Start with 4096 context tokens and 1024 output tokens; increase the output budget for longer pieces. Include your desired word count in the brief.

llama-cli -m AMKemma-3.1-Write-Q8_0.gguf -c 4096 -n 1024 --single-turn --reasoning off --temp 0.8 --top-p 1 --top-k 0 --min-p 0 --repeat-penalty 1 -p "Write a warm, informal travel diary about a rainy train journey. Around 500 words."

Two evaluation rounds

Both rounds test the same CE-best checkpoint against its untouched Gemma parent, on different frozen sets of 2,000 prompts.

	Round 1 — H2 / training-free ablator	Round 2 — FineWeb / Write family
Prompts	Original ablator TEST prompts	Muse briefs with style, content, outline and target length
Output budget	Fixed 1,024-token maximum	Per-document target-length budget
Sampling	Temperature 0.8, top-p 0.95, top-k 50	Temperature 0.8, top-p 1, top-k off
Comparisons	Fresh Gemma parent and human references	Fresh Gemma parent, humans and archived Write models
JSD units	Nats	Bits

Read each table within its own round. Dataset, prompting, budgets and sampling differ, so scores and lengths measure separate conditions. AMKemma writes shorter outputs than its parent in Round 1 and longer outputs in Round 2.

Evaluation — H2 TEST2000

Same 2000 frozen prompts for the untouched Gemma parent and AMKemma, alongside 2000 human reference texts. These measurements use the BF16 parent with the trained adapter, before GGUF quantization.

Sampler: temperature 0.8, top-p 0.95, top-k 50, repetition penalty 1, 1024-token cap, matched per-prompt seeds, native stops 106/1. All outputs are included.

Style — document means

Metric	Human	Gemma parent	AMKemma
Lexical diversity (unique lowercase tokens / tokens)	0.44899	0.41585	0.44999
Lexical density (non-stop alphabetic tokens / tokens)	0.44337	0.45670	0.43337
Long words (>6 characters / tokens)	0.19022	0.24708	0.18214
Token entropy (bits)	6.91822	6.89838	6.31738
Token-frequency burstiness (SD / mean)	1.71662	2.05725	1.37696
Vowel characters / characters	0.29663	0.30412	0.30332
Consonant characters / characters	0.47879	0.49033	0.48503
Punctuation characters / characters	0.03602	0.04567	0.03052
Length (NLTK tokens, including punctuation)	733.39250	637.51750	332.99650
Sentences / document	34.12800	30.41550	17.44400
Tokens / sentence	24.86669	23.64438	28.07701
Maximum parse depth	10.95950	10.30150	9.70050
Gunning Fog grade	12.98933	15.37583	12.21518
Linsear Write grade	11.59072	14.52612	11.01233

Lexical diversity, readability and long-word frequency move toward the human means. Outputs are shorter; entropy and burstiness also fall. Length-sensitive features reflect those different output lengths. NLTK length includes punctuation tokens. Full precision and paired 95% bootstrap intervals (1000 resamples) are in evaluation.json.

Corpus distance — lower means closer

Metric	Gemma parent	AMKemma	Human–human reference
L2-1	0.02906164	0.02214403	0.00807385
L2-2	0.01241105	0.01479894	0.00500227
L2-3	0.00661522	0.00430535	0.00394540
JSD-nats	0.10543271	0.05729762	0.08331861

L2 compares normalized corpus n-gram frequencies; JSD uses natural logs (nats). The human–human reference uses the archived VAL2 split protocol. Unigram and trigram distances improve; bigram distance increases. Fresh MMD is unavailable. FineWeb Muse 2000 results follow below.

Evaluation — FineWeb Muse 2000

Same 2,000 frozen Muse briefs and human pages used for the Write-family evaluation. Results use the original custom macro BLEU/chrF (0–1), JSD in bits, and the trained model before Q8 quantization.

Model	Scored n	BLEU ↑	chrF ↑	JSD bits ↓	Self-BLEU ↓	Repeat-start ↓
Gemma parent	1989	0.3887	0.6748	0.06744	0.2790	0.1735
AMKemma	2000	0.5545	0.7765	0.01846	0.3142	0.1630
Qwen baseline	2000	0.3959	0.7126	0.05386	0.2940	0.1620
FineWeb SFT	2000	0.5920	0.7941	0.02632	0.3213	0.2340
13-LoRA merger	2000	0.5750	0.7821	0.03194	0.3185	0.2530
ResidualBoost-S1	2000	0.5711	0.7754	0.02506	0.3174	0.2290

AMKemma improves overlap and corpus distance versus its Gemma parent. FineWeb SFT has higher overlap with the reference pages; AMKemma has the lowest unigram JSD in this comparison. Historical Qwen results use a separate model family and runtime.

All 4,000 fresh outputs are retained. The original scorer excludes 11 empty parent answers; AMKemma has zero. On the same 1,989 nonempty pairs, the BLEU gain is +0.1660 [0.1592, 0.1727] and chrF gain +0.1018 [0.0974, 0.1065] (paired 95% bootstrap intervals).

Style — document means

Feature	Human	Gemma parent	AMKemma
Lexical diversity	0.54876	0.56806	0.55170
Lexical density	0.47018	0.49061	0.46870
Long-word fraction (>6 characters)	0.19828	0.23058	0.19773
Token entropy (bits)	6.59491	6.43464	6.54348
Token-frequency burstiness (SD / mean)	1.29246	1.41038	1.28600
Vowel character fraction	0.30186	0.29921	0.30210
Consonant character fraction	0.48758	0.48628	0.48774
Punctuation character fraction	0.02887	0.03652	0.02771
Length (NLTK tokens, including punctuation)	504.32550	410.04300	452.43100
Sentences / document	22.23550	22.28850	21.75050
Tokens / sentence	24.23881	27.72495	25.25973
Maximum parse depth	10.02400	8.87100	9.91850
Gunning Fog grade	12.79795	15.08025	13.25387
Linsear Write grade	12.01844	12.88388	11.64308

AMKemma is closer to the human mean on 13 of 14 style features. Here its outputs are longer than the parent: 390.28 versus 336.90 whitespace words, with humans at 436.02. All 2,000 outputs per arm contribute to these means; features are measured at their generated lengths.

Protocol: temperature 0.8, top-p 1, top-k off, thinking off, 3,072-token input cap, original numeric target-length budgets, native Gemma stops 106/1 and matched per-prompt seeds. Budget hits: parent 805/2,000; AMKemma 906/2,000. Historical Qwen baseline uses a 2,048-token cap; 74 targets exceed it.

Full-precision results, corpus distances, paired intervals and common-pair coverage are in evaluation-fineweb.json. MMD/FID/MAUVE are unavailable; JMQ was skipped.

Dataset and collection

License

Apache 2.0. Based on Google Gemma 4. Attribution and modification details are in NOTICE.txt.

Downloads last month: 85

GGUF

Model size

5B params

Architecture

gemma4

Hardware compatibility

8-bit

Model tree for amkkk/AMKemma-3.1-Write

Base model

google/gemma-4-E2B

Finetuned

google/gemma-4-E2B-it

Adapter

(217)

this model

amkkk
/

AMKemma-3.1-Write

AMKemma - 3.1 Write

Abstract

Training

Files and usage

Two evaluation rounds

Evaluation — H2 TEST2000

Style — document means

Corpus distance — lower means closer

Evaluation — FineWeb Muse 2000

Style — document means

Dataset and collection

License

Model tree for amkkk/AMKemma-3.1-Write

Dataset used to train amkkk/AMKemma-3.1-Write