Facebook (胡嘉璽)
ai-engineering
Original source

Local LLM Model Naming Glossary

Context

Original post breaks down model names like Qwen3-30B-A3B-QAT-Instruct-Q4_K_M_MTP-NSFW for people new to local model deployment, explaining what each segment means.


Glossary

Parameter Size (參數量)

TermMeaning
BBillion parameters (e.g. 30B = 30 billion)
MMillion parameters (e.g. 125M = 125 million)
7B, 13B, 30B, 70BCommon model sizes; bigger = smarter but needs more VRAM

Architecture / Mixture of Experts (MoE)

TermMeaning
A3B“Active 3 Billion” — in MoE models, only a subset of parameters are active per token. 30B-A3B means 30B total params but only 3B active at inference time. Much faster than dense 30B.
MoEMixture of Experts — architecture that routes tokens to specialized sub-networks
DenseTraditional architecture where all parameters are used for every token

Quantization (量化)

TermMeaning
GGUFFile format for quantized models (used by llama.cpp, LM Studio, Ollama)
Q4_K_M4-bit quantization, K-quant method, Medium quality. Smaller file, slight quality loss.
Q4_K_S4-bit, K-quant, Small — more aggressive compression than M
Q5_K_M5-bit, K-quant, Medium — good balance of quality and size
Q6_K6-bit, K-quant — near-lossless for most use cases
Q8_08-bit quantization — very close to full precision
F16Full 16-bit floating point (no quantization)
BF16Brain Float 16 — Google’s format, same size as F16 but different precision distribution
MXFP4Microsoft’s mixed-precision 4-bit format
QATQuantization-Aware Training — model was trained knowing it would be quantized, so quality holds up better at low bits vs post-training quantization
AWQActivation-aware Weight Quantization — another quant method
GPTQPost-training quantization method (older, GPU-focused)
EXL2ExLlamaV2 quantization format, variable bits-per-weight
IQ4_XSImportance-matrix quantization, 4-bit extra-small — newer, quality-preserving

Fine-tuning Method (微調方法)

TermMeaning
InstructFine-tuned to follow instructions (chat format). Without this = base model that just does text completion.
ChatSimilar to Instruct — trained for conversational format
BaseRaw pretrained model, no instruction tuning
LoRALow-Rank Adaptation — lightweight fine-tuning that trains small adapter layers
DoRAWeight-Decomposed Low-Rank Adaptation — improved LoRA variant
QLoRALoRA applied to a quantized base model (saves VRAM during training)
SFTSupervised Fine-Tuning — training on curated instruction/response pairs
DPODirect Preference Optimization — alignment training using preference data
RLHFReinforcement Learning from Human Feedback — alignment method (used by ChatGPT)
MergeModel created by mathematically combining weights from multiple models
SLERP / TIES / DAREDifferent merge strategies

Inference Optimization (推理優化)

TermMeaning
MTPMulti-Token Prediction — model predicts multiple tokens at once for faster generation (speculative-style)
Speculative DecodingUsing a small draft model to propose tokens, verified by the main model
KV CacheKey-Value cache — stores past attention computations to avoid recomputation
Flash AttentionMemory-efficient attention algorithm
Context LengthMax tokens the model can process at once (e.g. 4K, 32K, 128K)
RoPERotary Position Embedding — enables context length extension

Content / Safety Tags

TermMeaning
NSFWNot Safe For Work — uncensored model, safety filters removed
UncensoredSame as NSFW — trained without refusal behavior
AbliteratedSafety training surgically removed via representation engineering
Censored / AlignedDefault — model will refuse harmful requests

Model Family Prefixes

TermMeaning
Qwen3Alibaba’s Qwen family, version 3
Llama-3Meta’s LLaMA family
Mistral / MixtralMistral AI models (Mixtral = MoE variant)
GemmaGoogle DeepMind’s open models
PhiMicrosoft’s small efficient models
DeepSeekDeepSeek AI models
Command-RCohere’s models

File / Deployment Terms

TermMeaning
safetensorsSafe file format for model weights (replaces pickle)
VRAMGPU memory needed to load the model
OffloadingPutting some layers on CPU RAM when VRAM isn’t enough (slower)
Context windowHow much text the model can “see” at once
TokensSub-word units; ~1 token ≈ 0.75 English words or ~0.5 Chinese characters

VRAM Quick Reference

Quant7B14B30B70B
Q4_K_M~5 GB~9 GB~19 GB~42 GB
Q8_0~8 GB~16 GB~34 GB~75 GB
F16~14 GB~28 GB~60 GB~140 GB

(Add ~1-2 GB overhead for KV cache at short context)


How to Read a Model Name

Example: Qwen3-30B-A3B-QAT-Instruct-Q4_K_M_MTP-NSFW

Qwen3       → Model family (Alibaba, version 3)
30B         → 30 billion total parameters
A3B         → Only 3B active per token (MoE)
QAT         → Quantization-Aware Training
Instruct    → Instruction-tuned (chat-ready)
Q4_K_M      → 4-bit K-quant Medium quantization
MTP         → Multi-Token Prediction enabled
NSFW        → Uncensored / no safety filters