Qwen 3.6 vs Gemma 4: testing on a local machine

# Comparing Qwen 3.6 and Gemma 4: Practical Testing on a Local Machine

The new local models Qwen 3.6 from Alibaba and Gemma 4 deliver impressive performance even on an average laptop. We tested them in real-world development conditions, setting up an environment for working with tools and code. The results confirm: both models integrate effectively into the workflow with proper configuration, but they require fine-tuning for limited resources.

Hardware Setup and Tool Selection

Testing was conducted on an Asus Tuf Gaming laptop with a discrete NVIDIA RTX 4070 graphics card (8 GB VRAM). Key constraints: limited video memory and the need to balance GPU acceleration with system resources. The following were selected for running the models:

• LM Studio — for loading and managing local models
• Zed IDE — as the primary development environment with an integrated AI agent

Key requirement: support for tools to work with the file system and terminal. Alternative CLI clients (OpenCode, Claude CLI, pi.dev) were ruled out due to:

• Failures when launching system prompts
• Incorrect shell detection (Bash/PowerShell)
• Issues with real-time file editing

Zed IDE showed stable performance after activating the agent via the "star" icon in the bottom-right corner (settings in the AI section).

Configuring LM Studio for Qwen 3.6

For Qwen 3.6 to work correctly in LM Studio, the Jinja template needed modification. Steps:

• In the MyModels section, select the Qwen 3.6 model
• Open settings (gear icon)
• In the Inference tab, replace the template with an optimized one

Critical changes:

• Adding {%- set enable_thinking = false %} at the beginning of the template to speed up responses
• Limiting GPU offload to 80-90% of total VRAM
• Setting context length ≤60K taking available RAM into account

Original template (abridged for example):

{%- set enable_thinking = false %}
{%- set image_count = namespace(value=0) %}
{%- macro render_content(content, do_vision_count, is_system_content=false) %}
    {%- if content is string %}
        {{- content }}
    {%- elif content is iterable and content is not mapping %}
        {%- for item in content %}
            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
                {%- if do_vision_count %}
                    {%- set image_count.value = image_count.value + 1 %}
                {%- endif %}
                {{- '<tool_call><tool_call><tool_call>' }}
            {%- elif 'text' in item %}
                {{- item.text }}
            {%- endif %}
        {%- endfor %}
    {%- endif %}
{%- endmacro %}

Exceeding 90% VRAM led to hallucinations and switching to Chinese/Spanish. The red warning in the LM Studio interface indicates a critical configuration error.

System Prompt and Speed Optimization

To shorten response lengths and save tokens, a specialized system prompt was used:

You are a professional React/NextJS developer. Respond briefly, without explanations. Use only necessary tokens. When working with files: first check the project structure, then make changes. Avoid generating code outside the current directory.

Optimization effects:

• Average response length reduced by 30%
• Generation speed increased by 15-20% due to disabling thinking mode
• Fewer errors when working with tools

Important: disabling thinking (enable_thinking = false) is critical for tasks that don't require multi-step reasoning. For complex algorithms, it's recommended to keep the mode enabled.

Performance Comparison

Testing was performed on tasks:

• Generating React components
• Debugging TypeScript
• Automating terminal commands

Qwen 3.6 (35B, Q4_K_M):

• Strength: accuracy when working with tools
• Weakness: high VRAM usage at context >48K
• Speed: 12-15 tokens/sec at 85% GPU offload

Gemma 4:

• Strength: stability at low VRAM
• Weakness: errors in multi-step tool calls
• Speed: 18-22 tokens/sec with the same settings

Both models handled prompts like this correctly:

Hi. You are a professional React/NextJS developer. In the current directory...

but Qwen 3.6 showed 25% fewer errors when working with the file system. Gemma 4 generated code faster but required manual fixes for tool calls.

Key Takeaways

• VRAM Balance: Never offload more than 90% of video memory to the GPU. For RTX 4070 (8 GB), 7 GB is optimal.
• Context Limit: 60K context length requires at least 16 GB RAM. On 32 GB systems, leave 4-5 GB for the OS.
• Tools: Qwen 3.6 is more stable with tools, but Gemma 4 is faster at code generation.
• Prompts: Disabling thinking speeds up responses but reduces quality for complex tasks.
• Environment: Zed IDE is preferable to CLI clients for integration with development tools.

— Editorial Team