Qwen3 8B
Alibaba Qwen
Switchable thinking mode: the same weights answer directly or reason step by step depending on the prompt. Long context for its size.
Excellent fit
- Weights
- 8.1 GB
- KV cache
- 1.1 GB
- Overhead
- 600 MB
- 9.8 GB of 23 GB usable VRAM.
- Chosen as the best quality that still fits a 32,768-token context (13 GB at that length).
- Hybrid thinking / non-thinking modes
- 128k context
- Very broad language coverage
- Thinking mode multiplies output tokens and latency
- Newer — less community tooling than Qwen2.5
qwen3:8bRunning Qwen3 8B on GeForce RTX 4090
| Quantization | Quality | Weights | KV cache | Total | ~tok/s | Fit |
|---|---|---|---|---|---|---|
| F16 | lossless | 15 GB | 1.1 GB | 17 GB | 48 | Good fit |
| Q8_0Pick | near-lossless | 8.1 GB | 1.1 GB | 9.8 GB | 89 | Excellent fit |
| Q6_K | near-lossless | 6.3 GB | 1.1 GB | 8.0 GB | 116 | Excellent fit |
| Q5_K_M | high | 5.4 GB | 1.1 GB | 7.1 GB | 134 | Excellent fit |
| Q4_K_M | balanced | 4.6 GB | 1.1 GB | 6.3 GB | 157 | Excellent fit |
| Q3_K_M | degraded | 3.7 GB | 1.1 GB | 5.5 GB | 194 | Excellent fit |
KV cache is sized at 8,192 tokens. Longer contexts cost proportionally more — the recommendation above reserves room for a working context.
You can fine-tune this here
- Base weights
- 4.3 GB
- Optimizer
- 650 MB
- Activations
- 900 MB
- Peak
- 7.4 GB
- 7.4 GB peak against 23 GB usable — room to raise batch size or sequence length.
- Base weights
- 15 GB
- Optimizer
- 650 MB
- Activations
- 900 MB
- Peak
- 18 GB
- 18 GB peak against 23 GB usable.
Where this model runs
VRAM 32 GB · F16 · 17 GB
VRAM 24 GB · Q8_0 · 9.8 GB
VRAM 16 GB · Q6_K · 8.0 GB
VRAM 16 GB · Q6_K · 8.0 GB
VRAM 24 GB · Q8_0 · 9.8 GB
VRAM 12 GB · Q5_K_M · 7.1 GB
VRAM 12 GB · Q5_K_M · 7.1 GB
VRAM 16 GB · Q6_K · 8.0 GB
VRAM 48 GB · F16 · 17 GB
VRAM 80 GB · F16 · 17 GB
Unified 128 GB · F16 · 17 GB
Unified 48 GB · F16 · 17 GB
Unified 24 GB · Q8_0 · 9.8 GB
Unified 192 GB · F16 · 17 GB
Unified 16 GB · Q6_K · 8.0 GB
VRAM 24 GB · Q8_0 · 9.8 GB
VRAM 16 GB · Q6_K · 8.0 GB
VRAM 0 MB
VRAM 0 MB
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 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 14B
14.8B · Apache 2.0
The Qwen3 mid-size. Long context and a reasoning mode inside a footprint a 24 GB card handles comfortably.
Llama 3.1 8B Instruct
8B · Llama 3.1 Community License
The most widely supported open model there is. If a tool, adapter or tutorial exists, it was written for this one first.
Catalogue figures come from each model’s published card. ModelLM has not independently measured them.