Model datasheet · Gemma 3
Gemma 3 27B
The conversationalist at scale: best-in-family prose, strong multilingual chat, real vision — and the receipts show it is NOT a coding or agent model (Aider 4.9%).
- Vendor
- Architecture
- Dense · 27B
- Context
- 131,072 tokens
- License
- Gemma Terms of Use
- Released
- 2025-03-12
- Vision
- Yes
→ GGUF (llama.cpp / LM Studio / Ollama) → MLX (Apple Silicon)
§1 Characteristics by quantization band
— Not yet rated · contributions welcome —
Q4–Q5 — 4–5 bit
17GB weights · e.g. Q4_0 , Q4_K_M , mlx-4bit + KV: 1GB at 8K · 2.9GB at 32K · 10.4GB at 128Kdata checked Aug 2026
- quant-degradation-community — community consensus @ Q4_0 · aggregated · 2026-08-30
Editorial derivation: Google's QAT Q4_0 preserves prose quality well.
data checked Aug 2026
- quant-degradation-community — community consensus @ Q4_0 · aggregated · 2026-08-30
Editorial derivation: small loss vs full band; QAT helps.
data checked Aug 2026
- quant-degradation-community — community consensus @ Q4_0 · aggregated · 2026-08-30
Editorial derivation: carried from full band.
— Not yet rated · contributions welcome —
FP16 — full precision (fp16/bf16)
54GB weights · e.g. bf16 + KV: 1GB at 8K · 2.9GB at 32K · 10.4GB at 128Kdata checked Aug 2026
- vendor-model-card — MATH (4-shot) = 50 @ bf16 · vendor-reported · 2026-08-30
GSM8K: 82.6.
data checked Aug 2026
- vendor-model-card — HumanEval = 48.8 @ bf16 · vendor-reported · 2026-08-30
MBPP: 65.6.
- aider-polyglot — percent correct (whole format) = 4.9 @ bf16 · aggregated · 2026-08-30
Near the bottom of the board — real-world code editing is far weaker than HumanEval suggests.
data checked Aug 2026
- vectara-hallucination — HHEM hallucination rate % (lower is better) = 7.4 @ bf16 · aggregated · 2026-08-30
Curiously worse than the 4B's 6.4%.
data checked Aug 2026
- community-consensus — community reputation (no public benchmark currently lists this model) @ bf16 · aggregated · 2026-08-30
Editorial score: the Gemma 3 prose reputation, strongest at 27B. No benchmark backs this cell.
§2 Known issues & what fixes them
Real-world code editing is dismal — 4.9% on the Aider polyglot board, near the bottom, despite a respectable-looking HumanEval. Snippet benchmarks flatter it; repository work exposes it.
Use a Qwen3 model for coding. Gemma's lane is chat, languages, writing, and vision.
Evidence · 2 sources · community-consensus
- Aider polyglot leaderboard (gemma-3-27b-it: 4.9%) (leaderboard, 2026-08-30)
- Model card (HumanEval 48.8 — the contrast) (vendor, 2026-08-30)
No native function-calling tokens or template — tool use rides on prompt-following, which works surprisingly well but breaks more than trained-in tool calling.
Give exact JSON schemas with examples in the system prompt — community benchmarks show it follows them decently.
For heavy MCP/agent work, pick a model with trained-in tool calling.
Evidence · 2 sources · community-consensus
- Discussion: Gemma 3 function-calling support (other, 2026-08-30)
- Gemma 3 prompt-based function-calling benchmarks (github, 2026-08-30)
Family-wide tendency to hallucinate names and specific factual details; HHEM (7.4%) is oddly worse than the 4B.
Ground factual work; treat unsourced names/dates with suspicion.
Evidence · 1 source · anecdotal
- Gemma-family factual-name hallucination reports (other, 2026-08-30)