Lesson 6: The Fine-Tune Zoo: Non-Quant Suffixes
Lesson 6: The Fine-Tune Zoo: Non-Quant Suffixes
Quant suffixes describe compression. This lesson covers the suffixes that describe content โ the fine-tunes and variants layered on top of the base weights. Same family (Qwen3.8-27B), very different behaviors.
The basics
| Suffix | Meaning | Notes |
|---|---|---|
base / no suffix | Raw pretrained weights | Good for fine-tuning and completions; bad for chat |
instruct, chat, it | Instruction/chat-tuned | The conversational model โ what you actually want for daily use |
thinking | Reasoning mode on by default | Qwen3.8 ships with thinking on, toggleable per request; uses more tokens |
vision, VL, omni | Multimodal | Qwen3.8-27B is natively vision-language โ in GGUF form you also need the mmproj-*.gguf projector file |
coder, math | Domain-tuned | e.g. Qwen3-Coder-30B-A3B-Instruct |
MTP | Multi-token prediction module | Separate small file (~1.4 GB) that drafts tokens for speculative decoding โ up to ~2x speed on supporting runtimes |
-2507, -2601 | Release date (YYMM) | Newer revision usually wins; check the model card for changes |
-1M, -128K, -32k | Context window | Long-context variants trade some efficiency for reach |
The uncensored family (the one everyone asks about)
uncensored, abliterated, OBLITERATED, Unleashed, Heretic, CRACK, AEON-ULTIMATE-UNCENSOREDโฆ dozens of names for the same idea: a fine-tune that strips the model's refusal behavior.
- Abliteration (the technical term) works by finding the "refusal direction" in the model's internal activations and deleting it โ no fine-tuning needed, so the rest of the weights stay intact.
- What changes: the model stops refusing requests it was trained to refuse. For creative writing, roleplay, horror, edgy humor โ great.
- What doesn't change: it's not "smarter", it's not free of safety features (it still has no guardrails โ that's a feature for RP, a liability for real work), and quality/format-following can dip slightly.
- Trust matters: these are community uploads. Prefer known-good publishers (huihui-ai, orcarouter, Blackfrost-AI) and check download counts + comments.
Speed and architecture words
- flash / DFlash2: FlashAttention or flash-style tuned variants โ faster attention and longer contexts. (Not a quality upgrade.)
- MoE naming:
30B-A3B= 30B total parameters, only ~3B active per token. Runs at the speed of a small model with more knowledge;A22B= 22B active, etc. - Distill: a smaller student model trained on a bigger teacher's outputs (e.g.
Fable-Distill). - Guard: safety classifiers (Qwen3Guard) โ not chat models.
- TTS / ASR / Embedding / Reranker: the Qwen3.8 family includes speech, embedding, and reranking models โ totally different tasks from chat; don't confuse their files with the LLM.
- Merge names (slerp, TIES, DARE): community model merges โ weight-space recipes blending several models. Can be brilliant or broken; trust the benchmarks, not the name.
- RP / story / NSFW: roleplay and creative-writing fine-tunes; usually derived from the uncensored family.
- Laser: sparsity-pruned (half the weights removed) โ faster, some quality loss.
- Turbo / Mini / Nano / Lite: size or speed tiers.
How fine-tunes change ability
Quantization changes precision; fine-tunes change behavior. A coder model writes better code but may lose general chat polish. An abliterated model writes freely but refuses nothing. A thinking model reasons more deeply but answers more slowly. When you see a suffix you don't know, ask three questions: Who made it? What did they change? Does the model card show benchmarks? If the answer to any is unclear, default to the official instruct release.
🧠 Knowledge Check
1. What does 'abliterated' mean?
2. 30B-A3B means:
3. In GGUF form, a vision-language model needs: