Known-issues sheets for local language models Rev 0.1.0 · reviewed every 7 days
llm-picker

Model datasheet · Llama 3.2

Llama 3.2 3B

Meta's edge-class model: built for summarization, rewriting, and on-device assistants, and honest about it. Runs on almost anything; don't ask it to be a scientist.

Vendor
Meta
Architecture
Dense · 3.2B
Context
131,072 tokens
License
Llama 3.2 Community License
Released
2024-09-25
Vision
No

→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)

§1 Characteristics by quantization band

Q2–Q3 — 2–3 bit

1.9GB weights · e.g. Q3_K_L + KV: 896MB at 8K · 3.5GB at 32K · 14GB at 128K

— Not yet rated · contributions welcome —

Q4–Q5 — 4–5 bit

2.1GB weights · e.g. Q4_K_M , mlx-4bit + KV: 896MB at 8K · 3.5GB at 32K · 14GB at 128K
Math & reasoning 4/10
data checked Aug 2026
  • quant-degradation-community — KLD/perplexity delta vs bf16 @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: small models lose proportionally more at 4-bit.

Instruction following 6/10
data checked Aug 2026
  • quant-degradation-community — community consensus @ Q4_K_M · aggregated · 2026-08-30

    Editorial derivation: IF is not quant-sensitive.

Tool calling / agentic 5/10
data checked Aug 2026

Q6–Q8 — 6–8 bit

2.7GB weights · e.g. Q6_K , Q8_0 + KV: 896MB at 8K · 3.5GB at 32K · 14GB at 128K

— Not yet rated · contributions welcome —

FP16 — full precision (fp16/bf16)

6.5GB weights · e.g. fp16 + KV: 896MB at 8K · 3.5GB at 32K · 14GB at 128K
Math & reasoning 5/10
data checked Aug 2026
  • vendor-model-card — MATH CoT = 48 @ bf16 · vendor-reported · 2026-08-30

    GSM8K: 77.7. No HumanEval/MBPP in the card, so coding is unrated.

Instruction following 6/10
data checked Aug 2026
Tool calling / agentic 6/10
data checked Aug 2026
Long context 5/10
data checked Aug 2026

§2 Known issues & what fixes them

ISSUE-01 moderate Q2–Q3 / Q4–Q5 / Q6–Q8 / FP16 · instruction-following, coding

A 3B model has a hard ceiling: answers get shallow, grammar and instruction-following slip on anything multi-step, and code quality is limited.

Fix
Prompt technique [partial]

Scope it to what Meta built it for — summarization, rewriting, extraction, short Q&A — one small task per prompt.

No fix
No known fix — choose a different model

For real coding or multi-step work, move up to an 8B+ model.

Evidence · 2 sources · community-consensus