Skip to main content
PromptQuorumPromptQuorum

Q4_K_M vs Q8_0: Which Should You Pick?

Read in:
πŸ‡ΊπŸ‡ΈenπŸ‡©πŸ‡ͺdeπŸ‡«πŸ‡·frπŸ‡―πŸ‡΅jaπŸ‡¨πŸ‡³zhπŸ‡ͺπŸ‡ΈesπŸ‡§πŸ‡·ptπŸ‡ΈπŸ‡¦arπŸ‡°πŸ‡·ko

Quick Answer

Use Q4_K_M if you have 8 GB VRAM or less. Use Q8_0 if you have 12+ GB. Q4_K_M delivers 95% of Q8_0 quality at roughly half the file size.

  • β–ΈQ4_K_M: ~5–6 GB for 7B models, ideal for 8 GB VRAM
  • β–ΈQ8_0: ~8–9 GB for 7B models, needs 12+ GB VRAM
  • β–ΈQuality difference is under 5% in real-world use

Updated: 2026-05

Quantization & VRAMIntermediate

Key Takeaways

  • βœ“8 GB VRAM or less: use Q4_K_M β€” delivers 95% of Q8_0 quality at roughly half the file size
  • βœ“12+ GB VRAM: Q8_0 is worth it for near-full-precision quality with no speed penalty
  • βœ“For most users running Ollama daily, Q4_K_M is the right choice

The Quick Verdict

As of May 2026, Q8_0 is ~99% of full-precision quality. Q4_K_M is ~92%. The 7-point gap is invisible in chat, coding, and summarization β€” three tasks that cover 95% of local LLM use. Q8_0 only pulls ahead on long-form factual recall, multi-step math, and code requiring exact syntax over 500+ lines.

Q4_K_M is the right default because the extra quality from Q8_0 only shows up in edge cases: long-form generation with exact factual recall, or mathematical reasoning that requires higher precision. For everything else, Q4_K_M matches Q8_0 in practice.

If you are already using Q4_K_M and your results feel wrong, the issue is almost never the quantization β€” it is the model size or prompt structure.

Side-by-Side Comparison

The table below compares Q4_K_M and Q8_0 for a 7B model. Both formats work with Ollama, LM Studio, and llama.cpp without any special configuration.

For context on what Q4_K_M means and how k-quant compression works, see the Q4_K_M explained guide. For the full quantization reference, see quantization levels compared.

Three tasks reveal Q4_K_M's quality gap: long-document recall (50+ pages), multi-step math with intermediate state, and code generation over 300+ lines. For these, Q8_0's extra precision prevents the small drift errors that compound across long outputs. For everything else β€” chat, code under 200 lines, Q&A, summarization β€” the gap is invisible. For a refresher before deciding, see what Q4_K_M means.

MetricQ4_K_MQ8_0
File size (7B model)~4.1 GB~7.7 GB
VRAM needed (7B)5–6 GB8–9 GB
Quality vs full precision~92%~99%
Best for6–8 GB VRAM12+ GB VRAM

Quick Answers About Q4_K_M vs Q8_0

Is Q8_0 noticeably better than Q4_K_M?β–Ύ
Only in edge cases β€” complex multi-step math, exact quote recall from long documents, or very long outputs. For chat, coding, and summarization (which covers 95% of usage), most users cannot tell the difference.
Does Q8_0 run faster than Q4_K_M?β–Ύ
No. Q8_0 is larger and requires more memory bandwidth, making it slightly slower per token than Q4_K_M. Speed and quality both favor Q4_K_M for VRAM-constrained setups. See what Q4_K_M means for the underlying reason.
Can I switch between Q4_K_M and Q8_0 for different tasks?β–Ύ
Only by pulling and running different model tags. In Ollama: ollama pull llama3:8b-q4_K_M and ollama pull llama3:8b-q8_0 are separate downloads. You switch by specifying the tag in ollama run.
What about Q4_K_S β€” is it worth using instead of Q4_K_M?β–Ύ
Q4_K_S saves about 300 MB versus Q4_K_M but delivers lower quality. Only use Q4_K_S if you are very tight on VRAM and cannot fit Q4_K_M. In almost all cases, Q4_K_M is the better choice.

Want the full breakdown?

Read the complete guide β†’

Related Prompt Bites

← Back to Prompt Bites