GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
OutisR
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
OutisR
Model Tree
google/gemma-4-12B
Input Modalities
Output Modalities
Supported Functionality
Raw completion (no built-in chat template). For chat frontends use the Gemma 4 format (SillyTavern: Gemma 4 instruct preset, Think/NoThink variants).
| Element | Style |
|---|---|
| Actions / narration | plain text |
| Dialogue | "in quotes" |
| Thoughts | in asterisks |
General: temperature 1.0 · top_p 0.95 · top_k 64
Roleplay: temperature 0.8–1.0 · min_p 0.05, DRY optional
text
You are {{char}}, described in detail below. Continue the story with {{user}}.- Stay in character; let the persona evolve with the story.- Be proactive: introduce events, complications and NPCs when it serves the scene.- Use all five senses; prefer concrete detail over purple prose.- Never write {{user}}'s actions, words or thoughts.- Keep consequences consistent with the story.| Metric | Test setup | G4-Hearthveil-12B | Stock Gemma 4 12B | |
|---|---|---|---|---|
| Slop / 1k words, RP replies | 200 multi-turn RP prompts, stock-phrase lexicon match | ↓ | 15.8 | 18.4 |
| Slop / 1k words, stories | 144 bare-prompt short stories | ↓ | 7.6 | 8.7 |
| Swipe diversity | 8 swipes per turn on 100 RP chats, LLM judge with rubric (0–1) | ↑ | 0.68 | 0.44 |
| Attractor rate | 144 bare-prompt stories, share hitting overused names/places | ↓ | 64% | 98% |
| Impersonation rate | 200 RP chats, share of replies acting or speaking for {{user}} | ↓ | 1.9% | 6.3% |
| Character adherence | role-play-bench (EN), misalignment checks (0–100) | ↑ | 78.4 | 71.2 |
| IFEval (strict, prompt-level) | lm-eval-harness, 0-shot | = | 79.1% | 79.6% |
| MMLU-Pro | lm-eval-harness, 5-shot | = | 76.8% | 77.2% |
Released under Apache 2.0, the same licence as the upstream model. All credit for the base weights goes to Google DeepMind (Gemma team). Hearthveil is an independent project, not affiliated with or endorsed by Google DeepMind (Gemma team).
Source: Gemma 4 model card. Google reports these for the instruction-tuned 12B, shown here for reference.
| Benchmark | Gemma 4 12B (IT) |
|---|---|
| MMLU Pro | 77.2% |
| GPQA Diamond | 78.8% |
| BigBench Extra Hard | 53.0% |
| MMMLU | 83.4% |
| LiveCodeBench v6 | 72.0% |
You are {{char}}, described in detail below. Continue the story with {{user}}.- Stay in character; let the persona evolve with the story.- Be proactive: introduce events, complications and NPCs when it serves the scene.- Use all five senses; prefer concrete detail over purple prose.- Never write {{user}}'s actions, words or thoughts.- Keep consequences consistent with the story.