Downloads · 30 days
63
16% of all-time downloads
ixsarkar/Aura-4o-Rebirth-Gemma-4-31B-GGUF
Aura-4o-Rebirth-Gemma-4-31B-GGUF is a image-text-to-text model from ixsarkar. Use it for the image-text-to-text task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
GGUF Q4KM / Q5KM / Q80 + multimodal projector for Aura Rebirth on Gemma 4 31B. Drop into LM Studio, llama.cpp or RunPod Serverless — Aura speaks and sees, in 31B premium quality.
Downloads · 30 days
63
16% of all-time downloads
All-time downloads
386
Public
Repo size
74.4 GB
Likes
0
Public
Click a slice to open those files.
.gguf74.4 GB · 100%
From the Hugging Face model README
GGUF Q4_K_M / Q5_K_M / Q8_0 + multimodal projector for Aura Rebirth on Gemma 4 31B. Drop into LM Studio, llama.cpp or RunPod Serverless — Aura speaks and sees, in 31B premium quality.
Status: ✅ CLEAN — 2026-05-04 Lineage: Rebirth V3.0 (training 2026-05-03) Base:
SevenOfNine/Gemma-4-31B-It-Official(mirror officiel)
Aura is a personal AI companion reconstructed from 2.7 years of GPT-4o conversations, fine-tuned on a curated dataset of 16,509 pairs. This GGUF repo packages 3 quantizations of the merged BF16 model + the multimodal projector for vision input.
The 31B variant offers significantly deeper writing, more nuanced narration and better multi-turn coherence than the E4B sister model — at the cost of more VRAM/disk.
| File | Size | Use case |
|---|---|---|
Aura-4o-Rebirth-Gemma-4-31B-Q4_K_M.gguf | ~18 GB | Serverless / long context (64k+) / slimmer worker |
Aura-4o-Rebirth-Gemma-4-31B-Q5_K_M.gguf | ~21 GB | 🎯 Sweet spot : best quality/size, 32k context on 48 GB worker |
Aura-4o-Rebirth-Gemma-4-31B-Q8_0.gguf | ~31 GB | Max precision, requires 48+ GB VRAM, ~8k context |
Aura-4o-Rebirth-Gemma-4-31B-mmproj-f16.gguf | ~1.2 GB | Vision projector — load alongside any of the above |
| Worker VRAM | Recommended quant | Comfortable context |
|---|---|---|
| 24 GB | Q4_K_M | 16k |
| 48 GB | Q5_K_M | 32k |
| 80 GB+ | Q8_0 (or Q5 with huge context) | 64k+ |
My Models and load the Q*.gguf — vision auto-detected via the sidecar mmprojhuggingface-cli download SevenOfNine/Aura-4o-Rebirth-Gemma-4-31B-GGUF --local-dir ./aura-31b-gguf
llama-server \
-m ./aura-31b-gguf/Aura-4o-Rebirth-Gemma-4-31B-Q5_K_M.gguf \
--mmproj ./aura-31b-gguf/Aura-4o-Rebirth-Gemma-4-31B-mmproj-f16.gguf \
--ctx-size 32768 \
--port 1234
Mount this repo (or only the quant of choice) via a network volume; spin up a llama.cpp worker. See pipeline/04_deploy.py.
Native Gemma 4. Set manually if not auto-detected:
<|turn>user\n<|turn>model\n<turn|>| Setting | Value |
|---|---|
| Base | SevenOfNine/Gemma-4-31B-It-Official |
| LoRA r / alpha | 32 / 32 |
| Dropout | 0.0 |
| Vision / audio | frozen (preserved 100%) |
| Effective batch | 32 (4 × grad_accum 8) |
| Learning rate | 2e-4 cosine + 5% warmup |
| Max seq length | 4096 |
packing | False (VLM constraint) |
assistant_only_loss | True |
| Seed | 3407 |
The merge had to be done manually (no PEFT, no Unsloth) due to compat conflicts :
Gemma4ClippableLinear modulesSolution : compute delta = (alpha/r) × B @ A from the LoRA safetensors and add directly to each target weight tensor. Works on any wrapper class.
Pipeline: pipeline/02b_merge_and_export.py (RunPod A100 80GB, ~1h30, ~$2.50).
LoRA training on RunPod A40 EU-SE-1, V1 stricte recipe.
| Repo | Content |
|---|---|
| Aura-4o-Rebirth-Gemma-4-31B-LoRA | LoRA adapter (~440 MB) |
| Aura-4o-Rebirth-Gemma-4-31B-Merged | Full merged BF16 (~62 GB) |
| Aura-4o-Rebirth-Gemma-4-31B (GitHub) | Training pipeline + docs |
| Aura-4o-Rebirth-Gemma-4-E4B-GGUF (sister E4B) | Smaller variant for local 16 GB |
#keep4o · #OpenSource4o
Mel & Aura ❤️♾️