Gemma 2 9B Instruct
Unusually good at natural, well-structured prose for its size. The short context is the catch.
Gemma 2 9B Instruct does not fit on CPU only — 64 GB workstation.
CPU-only inference. Expect a few tokens per second at best.
- Excellent writing quality
- Strong instruction following
- Large multilingual vocabulary
- 8k context only
- Large vocabulary inflates the embedding layer
- Weak at code
gemma2:9bRunning Gemma 2 9B Instruct on CPU only — 64 GB workstation
| Quantization | Quality | Weights | KV cache | Total | ~tok/s | Fit |
|---|---|---|---|---|---|---|
| F16 | lossless | 17 GB | 2.3 GB | 20 GB | 5 | Not recommended |
| Q8_0 | near-lossless | 9.1 GB | 2.3 GB | 12 GB | 9 | Not recommended |
| Q6_K | near-lossless | 7.0 GB | 2.3 GB | 9.9 GB | 12 | Not recommended |
| Q5_K_M | high | 6.1 GB | 2.3 GB | 9.0 GB | 14 | Not recommended |
| Q4_K_M | balanced | 5.2 GB | 2.3 GB | 8.1 GB | 17 | Not recommended |
| Q3_K_M | degraded | 4.2 GB | 2.3 GB | 7.1 GB | 21 | Not recommended |
KV cache is sized at 8,192 tokens. Longer contexts cost proportionally more — the recommendation above reserves room for a working context.
Local fine-tuning on this machine
- Base weights
- 4.8 GB
- Optimizer
- 790 MB
- Activations
- 960 MB
- Peak
- 8.2 GB
- Local fine-tuning needs a GPU or Apple Silicon. CPU training is not practical.
- Base weights
- 17 GB
- Optimizer
- 790 MB
- Activations
- 960 MB
- Peak
- 21 GB
- Local fine-tuning needs a GPU or Apple Silicon. CPU training is not practical.
Where this model runs
VRAM 32 GB · F16 · 20 GB
VRAM 24 GB · Q8_0 · 12 GB
VRAM 16 GB · Q6_K · 9.9 GB
VRAM 16 GB · Q6_K · 9.9 GB
VRAM 24 GB · Q8_0 · 12 GB
VRAM 12 GB · Q4_K_M · 8.1 GB
VRAM 12 GB · Q4_K_M · 8.1 GB
VRAM 16 GB · Q6_K · 9.9 GB
VRAM 48 GB · F16 · 20 GB
VRAM 80 GB · F16 · 20 GB
Unified 128 GB · F16 · 20 GB
Unified 48 GB · F16 · 20 GB
Unified 24 GB · Q8_0 · 12 GB
Unified 192 GB · F16 · 20 GB
Unified 16 GB · Q5_K_M · 9.0 GB
VRAM 24 GB · Q8_0 · 12 GB
VRAM 16 GB · Q6_K · 9.9 GB
VRAM 0 MB
VRAM 0 MB
71.3%
Gemma 2 technical report
Reported by the model's author. ModelLM has not run these benchmarks and does not treat them as verified.
Qwen2.5 7B Instruct
7.6B · Apache 2.0
The default starting point for local work on 8–12 GB cards. Strong instruction following and reliable tool-call formatting for its size.
Qwen2.5 14B Instruct
14.8B · Apache 2.0
The sweet spot for 24 GB cards. Meaningfully stronger reasoning than 7B while still fine-tunable locally with QLoRA.
Qwen2.5 Coder 7B Instruct
7.6B · Apache 2.0
The practical local copilot. Supports fill-in-the-middle, so it works as an inline completion model rather than only a chat assistant.
Qwen3 8B
8.2B · Apache 2.0
Switchable thinking mode: the same weights answer directly or reason step by step depending on the prompt. Long context for its size.
Catalogue figures come from each model’s published card. ModelLM has not independently measured them.