Qwen3.8 vs Qwen3.6 vs Gemma 4 on 24GB GPUs (10-minute read)

For local AI workloads on 24GB GPUs (e.g., RTX 4090), Qwen3.8-27B is evaluated as the strongest default choice among the three models. Under the same Q4_K_M quantization environment, it maintains decoding speeds similar to Qwen3.6-27B while passing all 12 coding tasks in a single seed and achieving high accuracy in reasoning and document QA.

Gemma 4 31B-it delivered the best results in structured tool calling (90/90) and vision tasks (19/20), but had drawbacks including high VRAM usage and slow initial token generation. Specifically, when using a 64K context, it required switching from an F16 K/V cache to Q8_0 to operate stably.

• Coding agents and repository editing: Qwen3.8-27B (12/12 pass@1)
• Strict tool calling: Gemma 4 31B-it (100% success in both single/multi-step)
• Vision-centric tasks: Gemma 4 31B-it (19/20)
• 64K context headroom: Qwen3.8 or Qwen3.6 (4.20 GiB headroom with F16 K/V)

If existing integrations are stable, it is advantageous to stick with Qwen3.6-27B, but for new deployments, Qwen3.8-27B offers a practical upgrade.