Gemma4 = hybrid or no?

[SmoothSailingMusic]

Howdy.
One thing that is throwing me off is that gemma4-26b-a4b models are not automatically detected as hybrid models.
When I use --force-hybrid, performance suffers greatly and is functionally unusable on my mac.
Turn 1 hermes prompt with "test"

That being said, I'm not filing it as a bug because it's possible I'm misunderstanding something.
Are the full hybrid benefits of gemma4 being utilized even when hybrid=false?
"INFO:rapid_mlx.model_auto_config:Auto-detected model family 'gemma[-_]?4' → tool_call_parser=gemma4, reasoning_parser=gemma4, is_hybrid=False, supports_spec_decode=True"
Or is this a case of gemma4 hybrid support not being available yet so when forcing hybrid it's utilizing an approach that is more qwen centred?
Cheers

[raullenchai]

Hi! Just to clarify, hybrid and MoE (Mixture of Experts) mean two different things in this context.

Gemma 4 26B is an MoE model, but Rapid-MLX does not classify it as a hybrid architecture. The auto-detected output you received is exactly right: is_moe=true and is_hybrid=false.

Please avoid using the --force-hybrid flag with Gemma 4. That flag is specifically intended for architectures like GatedDeltaNet or Mamba. Forcing it on Gemma 4 will bypass its native optimizations and significantly degrade your performance.

Rapid-MLX already handles Gemma 4's MoE routing and model-specific optimizations out of the box. You can simply run:

rapid-mlx serve gemma-4-26b-4bit

In short, your original automatic configuration was correct!

— Raullen

[SmoothSailingMusic]

Okay great ty for clarifying! Cheers

…

On Thu, Aug 20, 2026 at 2:38 PM Raullen Chai ***@***.***> wrote: Hi! Just to clarify, hybrid and MoE (Mixture of Experts) mean two different things in this context. Gemma 4 26B is an MoE model, but Rapid-MLX does not classify it as a hybrid architecture. The auto-detected output you received is exactly right: is_moe=true and is_hybrid=false. Please avoid using the --force-hybrid flag with Gemma 4. That flag is specifically intended for architectures like GatedDeltaNet or Mamba. Forcing it on Gemma 4 will bypass its native optimizations and significantly degrade your performance. Rapid-MLX already handles Gemma 4's MoE routing and model-specific optimizations out of the box. You can simply run: rapid-mlx serve gemma-4-26b-4bit In short, your original automatic configuration was correct! — Raullen — Reply to this email directly, view it on GitHub <#1195?email_source=notifications&email_token=AMF2EFWVL6F4YHVZQTKBHL35K5AS7A5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCOBQHE3TEMRSUZZGKYLTN5XKMYLVORUG64VFMV3GK3TUVRTG633UMVZF6Y3MNFRWW#discussioncomment-18097222>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMF2EFWJVJ2V4ESKQ23MQYD5K5AS7AVCNFSNUABJKJSXA33TNF2G64TZHMYTCNRWGE4DEMZVGE5UI2LTMN2XG43JN5XDWMJQGQ4DSNRYG6QXMAQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>