Best overallExcellent fit
Qwen2.5 14B Instruct
The sweet spot for 24 GB cards. Meaningfully stronger reasoning than 7B while still fine-tunable locally with QLoRA.
- Params
- 14.8B
- Quant
- Q6_K
- Memory
- 14 GB
Reference
Best local models for coding
Code models are judged on fill-in-the-middle support, repository-scale context and latency rather than general chat quality.
18 models fit23 GB usableEstimated
Top picks
Best on GeForce RTX 4090
Runner-upGood fit
Mistral Small 24B Instruct
Built deliberately for low latency on a single card — fewer layers, wider FFN. Apache 2.0 at a size that usually is not.
- Params
- 23.6B
- Quant
- Q4_K_M
- Memory
- 16 GB
ThirdGood fit
Phi-4 14B
Trained largely on curated synthetic data. Punches far above its size on reasoning and maths; the short context limits what you can do with it.
- Params
- 14.7B
- Quant
- Q8_0
- Memory
- 17 GB
Full results
18 models
Estimated
| # | Model | Params | Licence | Quantization | Memory | ~tok/s | Fit | Fine-tune |
|---|---|---|---|---|---|---|---|---|
| 1 | Qwen2.5 14B Instruct | 14.8B | Apache 2.0 | Q6_K | 14 GB | 64 | Excellent fit | Yes |
| 2 | Mistral Small 24B Instruct | 23.6B | Apache 2.0 | Q4_K_M | 16 GB | 55 | Good fit | Yes |
| 3 | Phi-4 14B | 14.7B | MIT | Q8_0 | 17 GB | 50 | Good fit | Yes |
| 4 | Qwen3 14B | 14.8B | Apache 2.0 | Q6_K | 13 GB | 64 | Excellent fit | Yes |
| 5 | DeepSeek-R1-Distill-Qwen-14B | 14.8B | MIT | Q6_K | 14 GB | 64 | Excellent fit | Yes |
| 6 | Qwen2.5 Coder 32B Instruct | 32.8B | Apache 2.0 | Q4_K_M | 21 GB | 39 | Tight fit | Yes |
| 7 | DeepSeek-Coder-V2-Lite Instruct | 15.7B | DeepSeek License | Q5_K_M | 13 GB | 458 | Excellent fit | Yes |
| 8 | StarCoder2 15B | 16B | BigCode OpenRAIL-M | Q8_0 | 17 GB | 46 | Good fit | Yes |
| 9 | Qwen2.5 32B Instruct | 32.8B | Apache 2.0 | Q4_K_M | 21 GB | 39 | Tight fit | Yes |
| 10 | Qwen2.5 Coder 7B Instruct | 7.6B | Apache 2.0 | F16 | 15 GB | 51 | Good fit | Yes |
| 11 | Llama 3.1 8B Instruct | 8B | Llama 3.1 Community License | Q8_0 | 9.5 GB | 92 | Excellent fit | Yes |
| 12 | Qwen3 8B | 8.2B | Apache 2.0 | Q8_0 | 9.8 GB | 89 | Excellent fit | Yes |
| 13 | Qwen2.5 7B Instruct | 7.6B | Apache 2.0 | F16 | 15 GB | 51 | Good fit | Yes |
| 14 | DeepSeek-R1-Distill-Qwen-32B | 32.8B | MIT | Q4_K_M | 21 GB | 39 | Tight fit | Yes |
| 15 | Code Llama 7B Instruct | 6.7B | Llama 2 Community License | Q8_0 | 11 GB | 109 | Excellent fit | Yes |
| 16 | Qwen2.5 72B Instruct | 72.7B | Qwen License | Q4_K_M | 45 GB | 18 | Runs with CPU offload | No |
| 17 | Llama 3.3 70B Instruct | 70.6B | Llama 3.3 Community License | Q4_K_M | 44 GB | 18 | Runs with CPU offload | No |
| 18 | Mixtral 8x7B Instruct | 46.7B | Apache 2.0 | Q4_K_M | 29 GB | 100 | Runs with CPU offload | No |
How this page is produced
The catalogue is filtered to models that declare this task, then ranked by capability — parameter count, published benchmarks, licence clarity — and derated by how comfortably each one fits the reference machine.
- · Ranked against an RTX 4090. Switch the hardware on any model page to size against your own machine.
- · Benchmark figures are publisher-reported, not measured by ModelLM.
- · 15 of these can also be fine-tuned on this hardware.
Check your own machine
ModelLM can read your actual GPU, VRAM and RAM and size every model against it.
Detect my hardwareRelated