recommend

Get top model recommendations for your hardware

Synopsis

llmfit recommend [OPTIONS]

Description

Provides curated recommendations of the best models for your system based on fit quality, use case, runtime, and other filters. This is optimized for programmatic use and defaults to JSON output.

Options

-n, --limit · integer · default: "5"

Number of recommendations to return.

--use-case · enum

Filter by use case category. Options:

• general - General-purpose models
• coding (alias: code) - Code generation and analysis
• reasoning (alias: reason) - Complex reasoning tasks
• chat - Conversational models
• multimodal (alias: vision) - Vision and multimodal
• embedding (alias: embed) - Text embeddings

--min-fit · enum · default: "marginal"

Filter by minimum fit level. Options:

• perfect - Only perfect fits
• good - Good or better
• marginal - Marginal or better (default)

--runtime · enum · default: "any"

Filter by inference runtime. Options:

• any - All runtimes (default)
• mlx - MLX only (Apple Silicon)
• llamacpp (aliases: llama.cpp, llama_cpp) - llama.cpp only

--json · boolean · default: "true"

Output as JSON. Default is true for this command.

--memory · string

Override GPU VRAM size (e.g., “32G”, “32000M”, “1.5T”).

--max-context · integer

Cap context length used for memory estimation (tokens). Must be >= 1.

Usage Examples

Basic Recommendations

# Get top 5 recommendations (default)
llmfit recommend

# Get top 10 recommendations
llmfit recommend -n 10

Filter by Use Case

# Best coding models
llmfit recommend --use-case coding -n 3

# Best reasoning models
llmfit recommend --use-case reasoning -n 5

# Best chat models
llmfit recommend --use-case chat

# Multimodal models
llmfit recommend --use-case multimodal

Filter by Fit Level

# Only perfect fits
llmfit recommend --min-fit perfect -n 10

# Good or better
llmfit recommend --min-fit good

# Include marginal fits (default)
llmfit recommend --min-fit marginal

Filter by Runtime

# MLX models only (Apple Silicon)
llmfit recommend --runtime mlx -n 5

# llama.cpp models only
llmfit recommend --runtime llamacpp -n 5

Combined Filters

# Top 3 perfect-fit coding models
llmfit recommend --use-case coding --min-fit perfect -n 3

# Top 5 good MLX models
llmfit recommend --runtime mlx --min-fit good -n 5

# Best reasoning models for 24GB VRAM
llmfit recommend --use-case reasoning --memory 24G

Human-Readable Output

# Disable JSON output
llmfit recommend --json false -n 10

Example Output

JSON Format (Default)

$ llmfit recommend -n 3

{
  "system": {
    "total_ram_gb": 64.0,
    "available_ram_gb": 58.24,
    "cpu_cores": 16,
    "cpu_name": "Apple M2 Max",
    "has_gpu": true,
    "gpu_vram_gb": 64.0,
    "unified_memory": true,
    "backend": "Metal"
  },
  "models": [
    {
      "name": "llama-3.3-70b",
      "provider": "Meta",
      "parameter_count": "70B",
      "params_b": 70.0,
      "context_length": 131072,
      "use_case": "general",
      "category": "General",
      "release_date": "2024-12-06",
      "fit_level": "perfect",
      "run_mode": "gpu",
      "score": 95.2,
      "estimated_tps": 42.5,
      "runtime": "MLX",
      "best_quant": "4bit",
      "memory_required_gb": 43.68,
      "utilization_pct": 68.2
    },
    {
      "name": "qwen-2.5-72b",
      "provider": "Alibaba",
      "parameter_count": "72B",
      "params_b": 72.0,
      "context_length": 32768,
      "use_case": "general",
      "category": "General",
      "release_date": "2024-09-19",
      "fit_level": "perfect",
      "run_mode": "gpu",
      "score": 94.8,
      "estimated_tps": 40.1,
      "runtime": "MLX",
      "best_quant": "4bit",
      "memory_required_gb": 45.79,
      "utilization_pct": 71.5
    },
    {
      "name": "deepseek-v3",
      "provider": "DeepSeek",
      "parameter_count": "671B",
      "params_b": 671.0,
      "context_length": 131072,
      "use_case": "reasoning",
      "category": "Reasoning",
      "release_date": "2024-12-26",
      "fit_level": "good",
      "run_mode": "moe_offload",
      "score": 92.1,
      "estimated_tps": 28.3,
      "runtime": "llamacpp",
      "best_quant": "Q4_K_M",
      "memory_required_gb": 57.42,
      "utilization_pct": 89.7
    }
  ]
}

Filtered by Use Case

$ llmfit recommend --use-case coding -n 3

{
  "system": { ... },
  "models": [
    {
      "name": "qwen-2.5-coder-32b",
      "provider": "Alibaba",
      "parameter_count": "32B",
      "use_case": "coding",
      "category": "Coding",
      "fit_level": "perfect",
      "score": 91.2,
      "estimated_tps": 68.2
    },
    {
      "name": "codestral-25.01",
      "provider": "Mistral",
      "parameter_count": "22B",
      "use_case": "coding",
      "category": "Coding",
      "fit_level": "perfect",
      "score": 89.5,
      "estimated_tps": 85.1
    },
    {
      "name": "deepseek-coder-v2",
      "provider": "DeepSeek",
      "parameter_count": "236B",
      "use_case": "coding",
      "category": "Coding",
      "fit_level": "good",
      "score": 87.3,
      "estimated_tps": 32.7
    }
  ]
}

Human-Readable Format

$ llmfit recommend --json false -n 5

╭─ System Hardware ──────────────────────────────────────────╮
│  RAM:  64.0 GB total (58.2 GB available)                  │
│  CPU:  16 cores (Apple M2 Max)                            │
│  GPU:  Metal - Apple M2 Max (64.0 GB, unified memory)     │
╰────────────────────────────────────────────────────────────╯

╭─────────────┬──────────────────────┬───────────┬──────┬───────┬──────────────┬─────────┬────────────┬─────────┬────────┬─────────╮
│ Status      │ Model                │ Provider  │ Size │ Score │ tok/s est.   │ Quant   │ Runtime    │ Mode    │ Mem %  │ Context │
├─────────────┼──────────────────────┼───────────┼──────┼───────┼──────────────┼─────────┼────────────┼─────────┼────────┼─────────┤
│ ✓ Perfect   │ llama-3.3-70b        │ Meta      │ 70B  │ 95    │ 42.5         │ 4bit    │ MLX        │ GPU     │ 68.2%  │ 128k    │
│ ✓ Perfect   │ qwen-2.5-72b         │ Alibaba   │ 72B  │ 95    │ 40.1         │ 4bit    │ MLX        │ GPU     │ 71.5%  │ 32k     │
│ ✓ Good      │ deepseek-v3          │ DeepSeek  │ 671B │ 92    │ 28.3         │ Q4_K_M  │ llama.cpp  │ GPU     │ 89.7%  │ 128k    │
│ ✓ Perfect   │ qwen-2.5-coder-32b   │ Alibaba   │ 32B  │ 91    │ 68.2         │ 4bit    │ MLX        │ GPU     │ 52.3%  │ 128k    │
│ ✓ Perfect   │ llama-3.1-70b        │ Meta      │ 70B  │ 91    │ 42.5         │ 4bit    │ MLX        │ GPU     │ 68.2%  │ 128k    │
╰─────────────┴──────────────────────┴───────────┴──────┴───────┴──────────────┴─────────┴────────────┴─────────┴────────┴─────────╯

Filtering Behavior

Backend Compatibility

• MLX-only models are automatically hidden on non-Apple Silicon systems
• CUDA/ROCm models require appropriate GPU drivers

Fit Level Filtering

• --min-fit perfect: Only models with Perfect fit level
• --min-fit good: Perfect and Good fit levels
• --min-fit marginal: Perfect, Good, and Marginal (default)
• Too Tight models are excluded by default

Use Case Categories

• General: Versatile models for various tasks
• Coding: Code generation, analysis, completion
• Reasoning: Complex logical and mathematical reasoning
• Chat: Conversational and instruction-following
• Multimodal: Vision and image understanding
• Embedding: Text embedding generation

Use Cases

CI/CD Integration

#!/bin/bash
# Select best model for deployment
BEST_MODEL=$(llmfit recommend -n 1 --min-fit perfect | jq -r '.models[0].name')
echo "Deploying $BEST_MODEL"

Auto-Configuration

# Get best coding model for development environment
CODER=$(llmfit recommend --use-case coding -n 1 | jq -r '.models[0].name')
ollama pull $CODER

Hardware Selection

# Test different VRAM configurations
for vram in 24 32 48 64 80; do
  echo "== ${vram}GB VRAM =="
  llmfit recommend --memory "${vram}G" --min-fit perfect -n 3 | jq '.models[].name'
done

Related Commands

• fit - See all compatible models
• info - Get detailed model information
• system - Verify detected hardware
• search - Search for specific models