| # | Modelo | Params | MTEB Retrieval | LegalBench | MMLU-Law | Licencia | Servidor (A10G) | Local (M2 Pro) |
|---|---|---|---|---|---|---|---|---|
⭐ |
Mistral-7B-Instruct-v0.3
TOP PICK
mistralai · Context: 32k tokens
|
7.2B |
67.4%
|
71.8% |
74.2% |
Apache 2.0 | ✓ fp16 | ✓ fp16 |
2 |
Qwen · Context: 128k tokens
|
14.7B |
68.9%
|
73.5% |
77.1% |
Apache 2.0 | Q4 only necesita quantización | Q4 only ~10 GB Q4_K_M |
3 |
meta-llama · Context: 128k tokens
|
8.0B |
65.1%
|
68.3% |
72.9% |
Llama 3.1 | ✓ fp16 | Q4 only ~5.5 GB Q4_K_M |
4 |
google · Context: 8k tokens
|
9.2B |
66.7%
|
64.1% |
73.5% |
Gemma | ✓ fp16 | Q4 only context 8k — limitado para RAG |
5 |
Qwen · Context: 128k tokens
|
7.6B |
63.2%
|
66.8% |
72.1% |
Apache 2.0 | ✓ fp16 | ✓ fp16 |
6 |
meta-llama · Context: 128k tokens
|
70.6B |
72.4%
|
78.9% |
82.3% |
Llama 3.1 | ✗ demasiado grande necesita multi-GPU | ✗ demasiado grande |
7 |
mistralai · MoE · Context: 32k tokens
|
46.7B |
69.8%
|
70.2% |
76.4% |
Apache 2.0 | Q4 only ~27 GB Q4 — ajustado | ✗ demasiado grande |
# Instalar vLLM pip install vllm # Servir el modelo (fp16 nativo) vllm serve mistralai/Mistral-7B-Instruct-v0.3 \ --dtype float16 \ --max-model-len 32768 \ --gpu-memory-utilization 0.90 \ --port 8000 # Test rápido curl http://localhost:8000/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"model":"mistralai/Mistral-7B-Instruct-v0.3", "messages":[{"role":"user","content": "Extrae las partes firmantes del contrato: ..."}]}'
# Instalar ollama (macOS) brew install ollama # Descargar y servir Mistral 7B ollama run mistral:7b-instruct-v0.3 # O con llama.cpp directo (fp16) huggingface-cli download mistralai/Mistral-7B-Instruct-v0.3 \ --local-dir ./models/mistral-7b llama-server -m ./models/mistral-7b/model.safetensors \ -c 32768 --n-gpu-layers -1 --port 8080 # Estima: ~14.5 GB RAM unificada en fp16