Compare LLMs side by side
Put 2–4 models side by side — specs, capabilities and per-task fit — to see which one is better for which use, from live Hugging Face data.
Models
2 / 4Qwen2.5-7B-Instruct Qwen/Qwen2.5-7B-Instruct | Qwen2.5-Coder-7B-Instruct Qwen/Qwen2.5-Coder-7B-Instruct | |
|---|---|---|
| Identity & popularity | ||
| Parametersⓘ | — | — |
| MoE (active / total)ⓘ | — | — |
| Contextⓘ | — | — |
| Licenseⓘ | — | — |
| Downloads (30d)ⓘ | — | — |
| Likesⓘ | — | — |
| Updatedⓘ | — | — |
| Base modelⓘ | — | — |
| Architecture & memory | ||
| Precisionⓘ | — | — |
| Layersⓘ | — | — |
| Hidden sizeⓘ | — | — |
| Attn / KV headsⓘ | — | — |
| Head dimⓘ | — | — |
| Vocabⓘ | — | — |
| Weights (native)ⓘ | — | — |
| KV / tokenⓘ | — | — |
| VRAM @ 8k · 1 userⓘ | — | — |
| Capabilities | ||
| Instruction-tunedⓘ | — | — |
| Codeⓘ | — | — |
| Mathⓘ | — | — |
| Reasoningⓘ | — | — |
| Visionⓘ | — | — |
| Multilingualⓘ | — | — |
| Embeddingsⓘ | — | — |
| Task fitscore 0–100 · best highlighted | ||
| Chatbot / assistantⓘ | — | — |
| Code generationⓘ | — | — |
| RAG / doc Q&Aⓘ | — | — |
| Long-document summarizationⓘ | — | — |
| Agents / tool useⓘ | — | — |
| Reasoning & mathⓘ | — | — |
| Multilingualⓘ | — | — |
| Vision / multimodalⓘ | — | — |
| Embeddings / searchⓘ | — | — |
| On-device / edgeⓘ | — | — |
| High-throughput servingⓘ | — | — |
What the comparison measures
Put two to four models from the Hugging Face Hub side by side. Each column is resolved the same way as on the other pages: parameters, experts, maximum context, licence, downloads, likes, last update and base model; the precision the weights ship in and what they weigh; the KV cache per token at FP16; and the GPU memory needed for 8k tokens and one user.
Below that come the capabilities the Hub declares — instruct, code, math, reasoning, vision, multilingual, embedding — and the task-fit score for all 11 tasks. The best value is highlighted for context, downloads, likes and each task score; a row where every model ties stays plain.
Questions
- Is the VRAM figure the same as the calculator's?
- It comes from the same engine at a fixed workload: 8k context, one user, FP16 KV cache and vLLM's overhead on one GPU. Open the model in the VRAM calculator to size it for your own context and number of users.
- Why does a quantized model sometimes show BF16?
- When the precision cannot be detected from the quantization config, the repository name, torch_dtype or the safetensors header, the comparison assumes BF16. If the weight size looks too large, check the repository.
- Which model is better for my task?
- The task rows score how well each model is built for it — instruction tuning, context, size, declared capabilities. They cannot measure answer quality, so use them to shortlist, then try the finalists on your own prompts.
- How many models can I compare?
- Up to four at once. The selection lives in the page address, so a link opens the same comparison for someone else.