Qwen 3.6 27B: The Ideal Balance for Local AI Development

By Piotr Migdał | June 29, 2026

For a long time, I found local LLMs lacking. However, I've finally encountered a model that feels like a genuine leap toward general intelligence for local use.

The Qwen 3.6 series offers two primary paths:

1. Qwen 3.6 35B A3B: A Mixture-of-Experts (MoE) architecture.
2. Qwen 3.6 27B: A dense model that is slower but significantly more capable.

🚀 Testing the Capabilities

It is widely noted that the 27B variant "punches above its weight class." To verify this, I looked at a few different benchmarks and personal tests.

The "Smoke Tests"

Simon Willison often uses the "penguins on a bicycle" prompt as a litmus test for model reasoning. Both the 35B A3B and the 27B passed, proving that capabilities once reserved for prohibitively expensive models like GPT-4.5 (now legacy) are now available locally.

Creative & Technical Logic

I challenged the model to compose an 8-line poem blending quantum physics with Zouk dance. The result was impressive; the model didn't just rhyme, but actually deliberated on the quantum terminology.

When it came to coding, I used OpenCode to request a hexagonal minesweeper built with pnpm.

• Qwen 3.6 27B: Nailed it on the first attempt, providing a complete, functional Node package.
• Qwen 3.6 35B A3B: While faster, it failed the specific instruction to create a package, instead dumping everything into a single index.html.

"Real work isn't just about minesweeper clones."

To test a "real-world" scenario, a friend (Maciej Cielecki) provided a prompt for a landing page. The 27B model produced a reactive page with polished defaults from a single, brief prompt.

---

🛠️ Local Deployment Guide

Setting up local models is now remarkably streamlined. I strongly suggest using llama.cpp over Ollama (due to ethical concerns).

Setup Checklist

• Visit Hugging Face for quantized weights.
• Select a provider (e.g., unsloth or bartowski).
• Choose a quantization level (8-bit is the "sweet spot" for quality vs. size).
• Download the GGUF file.

I recommend the unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 version, which supports Multi-Token Prediction (MTP).

Execution Command

Run the following in your terminal to start the server:

llama-server -hf unsloth/Qwen3.6-27B-MTP-GGUF:Q8_0 \
  --spec-type draft-mtp \
  -ngl 999 \
  -fa on \
  -c 65536 \
  --port 8080

Parameter Breakdown:

• -hf: Pulls directly from Hugging Face (or use -m for a local path).
• --spec-type draft-mtp: Uses a draft model to predict tokens, increasing speed.
• -ngl 999: Offloads all layers to the GPU.
• -fa on: Enables Flash Attention.
• -c 65536: Sets context to 64k (though the model supports up to 256k).
• --port 8080: Pins the API port.

Integration with OpenCode

To connect this to your agent, update your ~/.config/opencode/opencode.jsonc:

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "llama": {
      "name": "llama.cpp (local)",
      "npm": "@ai-sdk/openai-compatible",
      "options": {
        "baseURL": "http://127.0.0.1:8080/v1",
        "apiKey": "local"
      },
      "models": {
        "qwen3.6-27b": { "name": "Qwen3.6-27B Q8 +MTP" }
      }
    }
  },
  "model": "llama/qwen3.6-27b"
}

Note: If you prefer a terminal-only experience, use llama-cli instead of llama-server.

---

📊 Performance Analysis

I benchmarked these models on a Macbook Max M5 (128 GB RAM).

Key Takeaways:

• Efficiency: llama.cpp outperformed mlx-lm on the M5, utilizing 95% of the GPU.
• Accessibility: Both variants fit within 48 GB of shared RAM. A 4-bit quantization would fit on a 32 GB machine ($\text{RAM} < 18\text{ GB}$).
• Nvidia Performance: On an RTX 5090 (Q6_K quantization), users report $\approx 50\text{ tok/s}$ with a 123k context using LM Studio.

🏁 Final Verdict

While the 35B A3B is roughly $3\times$ faster, the 27B Dense model is the superior choice for development. I would rather wait longer for code that is correct the first time than generate fast, mediocre code.

According to Artificial Analysis, the intelligence scaling looks like this: