Practical guide for running AI models locally on M4 MacBook: 24GB of memory can run 128K context smoothly

Core conclusion

Local AI models are no longer toys. In May 2026, a developer (Johanna Larsson) successfully ran the Qwen 3.5-9B Q4 quantization model on an M4 MacBook Pro (24GB memory), achieving 40 tokens/s inference speed, 128K context window and available Tool Use capabilities. This practice shows that owning a mid-range MacBook is enough to build an AI development assistant that does not rely on cloud services.

Key Points

• Event source: Hacker News hot post (229 likes), posted by Johanna Larsson on jola.dev
• Core Findings: Qwen 3.5-9B Q4_K_S quantified as the only model that combines speed, context width, and tool availability on a 24GB M4 MacBook
• Applicable people: Content creators, independent developers, small teams, who need local AI assistance but don’t want to pay for cloud APIs
• Cost comparison: one-time hardware investment (M4 MacBook ≈ ¥12,000), no subsequent API fees

Background: The "Impossible Triangle" of local AI

For a long time, running AI models locally has faced an "impossible triangle":

Best solution tested: Qwen 3.5-9B + LM Studio

After many attempts (Qwen 3.6 Q3, GPT-OSS 20B, Devstral Small 24B, Gemma 4B), the author finally confirmed that Qwen 3.5-9B Q4_K_S is currently the most balanced choice on the 24GB M4 MacBook.

Recommended configuration

// LM Studio
temperature: 0.6
top_p: 0.95
top_k: 20
min_p: 0.0
presence_penalty: 0.0
repetition_penalty: 1.0

// Thinking
// → → Prompt Template
// {% raw %}{% set enable_thinking = true %}{% endraw %}

Pi

{
  "providers": {
    "lmstudio": {
      "baseUrl": "http://localhost:1234/v1",
      "api": "openai-completions",
      "apiKey": "lm-studio",
      "models": [
        {
          "id": "qwen3.5-9b@q4_k_s",
          "reasoning": true,
          "compat": {
            "thinkingFormat": "qwen-chat-template"
          }
        }
      ]
    }
  }
}

OpenCode

{
  "provider": {
    "lmstudio": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LM Studio (local)",
      "options": {
        "baseURL": "http://127.0.0.1:1234/v1"
      },
      "models": {
        "qwen3.5-9b@q4_k_s": {
          "name": "Qwen 3.5 9B Q4_K_S",
          "tools": true,
          "context_length": 131072,
          "max_tokens": 32768
        }
      }
    }
  },
  "model": "lmstudio/qwen3.5-9b@q4_k_s"
}

vs SOTA

、、****。

****。

「 SOTA ，，。，——。」

Hacker News 「Task Paralysis and AI」（221 ）AI 。

• ****、、
• ****、、
• ****、、

1. LM Studio
2. Hugging Face Qwen 3.5-9B Q4_K_S
3. ， Thinking
4. API （LM Studio ）
5. AI （Pi OpenCode）

Apple Silicon （M4 Ultra / M5）（GGUF Q3/Q2），。，「 80% + API 20% 」，。

LM Studio、OpenCode、Qwen、Hugging Face、Claude、ChatGPT、Ollama、llama.cpp

Internal link guidance

• Want to try it yourself? See the full tutorial: How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial
• Real case: AI developers can make money by using local models to build automated workflows: Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo
• Advanced reading: DeepSeek V4 output price is only 1/36 of GPT-5.5: Token economic reconstruction in progress

Reference material

• Running local models on an M4 with 24GB memory
• Hacker News
• LM Studio
• Qwen 3.5-9B Q4_K_S on Hugging Face