Skip to content
TheChatArchiveThe Chat Archive
Perplexitysonar-reasoning· PerplexityHugging Face LMArena Search Arena 24K

Find the 5 best local LLMs available in March 2025 that would work on an M1 Macbook with…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with gpt-4o-search-preview’s answer.
  1. User

    Find the 5 best local LLMs available in March 2025 that would work on an M1 Macbook with 16GB of RAM, and compare them to the best available models anywhere in March 2023.

  2. Perplexity Assistant

    Top 5 Local LLMs for M1 MacBooks (16GB RAM) in March 2025

    1. Mistral Small 3 (24B)

    • Optimized for low latency (150 tokens/sec) and designed for real-time applications[6].
    • Quantized versions (e.g., Q4_K_M) fit within 16GB RAM when using tools like LM Studio or Ollama[1][7].
    • License: Apache 2.0 (open-source)[6].

    2. DeepSeek R1 (Quantized)

    • State-of-the-art reasoning but requires 4-bit quantization to run on 16GB RAM[3][9].
    • Performance: Slower than non-quantized models but feasible on M1 Macs with unified memory[3].

    3. Gemma 2 (9B/27B)

    • Google’s open-source alternative to Gemini, with 8,200-token context windows[6].
    • Use Case: Ideal for cost-sensitive deployments needing multimodal capabilities[6].

    4. LLaMa 3 (7B)

    • Meta’s model optimized for code generation and general-purpose tasks[8].
    • RAM Efficiency: Quantized versions (e.g., Q4_K_M) run smoothly on 16GB systems[5][8].

    5. Mistral 7B Instruct

    • Balanced performance for chat and coding tasks, widely supported in tools like LM Studio[1][7].
    • Quantization: Q4_K_M versions reduce memory footprint[1].

    Comparison: March 2025 vs. March 2023 Models

    Metric2025 Models2023 Models
    Top Open-SourceMistral Small 3, DeepSeek R1Mistral 7B, LLaMa 2
    Latency150+ tokens/sec (Mistral Small 3)[6]~50 tokens/sec (Mistral 7B)[5]
    Quantization SupportWidespread (4-bit/8-bit)[3][9]Limited (mostly 4-bit)[5]
    Context WindowUp to 128k tokens (Command R+)[6]~4k tokens (Claude 1.0)[6]
    Hardware OptimizationApple Metal GPU support[7][9]Primarily Nvidia GPU-focused[1][5]

    Key Advancements in 2025

    • Unified Memory: M1/M2/M3 chips enable efficient model loading without VRAM constraints[3][9].
    • Tooling: LM Studio and Ollama simplify deployment, while mlx/LLaMa.cpp optimize performance[1][4][7].
    • Quantization: 4-bit models (e.g., DeepSeek R1) enable larger models on consumer hardware[3][9].

    2023 Limitations:

    • GPT-3.5/Claude 1.0: Proprietary and cloud-dependent, with no local options[6].
    • Mistral 7B: Required 16GB+ RAM for non-quantized versions, limiting accessibility[5].

    For M1 MacBooks, Mistral Small 3 and LLaMa 3 offer the best balance of performance and efficiency, while DeepSeek R1 pushes the boundaries of local AI capabilities[3][6][8].

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.