Abiray/Artemis-31B-v1.2-GGUF

Artemis-31B-v1.2 GGUF

Static GGUF quantizations of TheDrummer/Artemis-31B-v1.2.

Artemis-31B-v1.2 is a fine-tuned, unaligned creative writing and roleplay model built on Google's Gemma 4 31B base architecture. It is specifically calibrated for literary depth, psychological nuance, long-horizon scene continuity, and multi-turn roleplay.

---

Available Files & VRAM Recommendations

Filename	Size	Recommended VRAM	Description & Use Case
Artemis-31B-v1.2-Q3_K_M.gguf	15.3 GB	18 – 20 GB	Low-memory footprint; runs comfortably on constrained setups.
Artemis-31B-v1.2-Q4_K_S.gguf	17.8 GB	20 – 22 GB	Lean 4-bit quant; fits tighter single-GPU configurations.
Artemis-31B-v1.2-Q4_K_M.gguf	18.7 GB	24 GB	Community Sweet Spot. Balances quality retention and context headroom.
Artemis-31B-v1.2-Q5_K_M.gguf	21.8 GB	24 – 32 GB	Near-lossless precision; exceptional for descriptive prose and dialogue fidelity.
Artemis-31B-v1.2-Q6_K.gguf	25.2 GB	32 – 40 GB	High-tier quant; indistinguishable from native 16-bit float.
Artemis-31B-v1.2-Q8_0.gguf	32.6 GB	40 – 48 GB	Reference baseline quant with near-zero perplexity loss.

---

Prompt Formats

Artemis-31B uses the Gemma 4 Chat Template and natively supports both Thinking (Reasoning Scratchpad) and Direct Storytelling operational modes.

1. Thinking Mode (Recommended for Complex Plots)

Trigger the reasoning planner by embedding <|think|> into the prompt. The model will plan character motivations, spatial distance, and scene tone in a scratchpad before generating the visible narrative.

<start_of_turn>user
<|think|>
Write a scene where two rival mercenaries negotiate a temporary truce in an abandoned chapel during a midnight thunderstorm.<end_of_turn>
<start_of_turn>model
<|channel>thought
The scene demands tension and heavy atmospheric detail. Both characters should remain guarded, tracking the other's weapon hand.
<|channel>call
Rain rattled against the leaded glass like loose teeth...

2. Standard Non-Thinking Mode

For traditional turn-by-turn dialogue without reasoning tokens:

<start_of_turn>user
{{user_prompt}}<end_of_turn>
<start_of_turn>model
{{model_response}}<end_of_turn>

---

Sampler Guidance (Anti-"Dash Spiral")

Because Artemis possesses a wide, expressive vocabulary distribution, standard greedy samplers can sometimes trigger repetitive loops or excessive em-dashes (—). The community consensus recommends pairing the model with Min-P:

Recommended Daily Driver (SillyTavern / KoboldCpp)

• Temperature: 0.95 – 1.05
• Min-P: 0.08
• Top-P: Disabled (1.0)
• Repetition Penalty: 1.06 – 1.08
• Repetition Penalty Range: 2048 tokens
• Presence / Frequency Penalty: 0.00

---

Quickstart Guide

Running via llama.cpp

# Serve as a local OpenAI-compatible API endpoint
llama-server \
  -hf Abiray/Artemis-31B-v1.2-GGUF:Q4_K_M \
  -c 8192 \
  -ngl 99 \
  --temp 0.95 \
  --min-p 0.08 \
  --repeat-penalty 1.06 \
  --port 8080

---

Credits & Acknowledgments

• Original Model: TheDrummer for creating and fine-tuning Artemis-31B-v1.2.
• Base Architecture: Google DeepMind's Gemma 4 31B.
• Quantization Engine: llama.cpp by Georgi Gerganov and contributors.