Model comparison · Updated November 18, 2026
Qwen 3.8 27B vs Gemma 4 31B
Qwen 3.8 27B Abliterated is the safer default on a 24 GB GPU, and Gemma 4 31B Heretic is the alternative.
Side by side
| Qwen 3.8 27B Abliterated | Gemma 4 31B Heretic | |
|---|---|---|
| Total parametersFrom Hugging Face | 27.8B | 31.3B |
| Input | text | text |
| Input pricePer 1M tokens, lowest host | $1.60 | Not hosted |
| Output pricePer 1M tokens, lowest host | $12.00 | Not hosted |
| Blended price3 input to 1 output | $4.20 | Not hosted |
| Size at 4-bit | ~17 GB at 4-bit | ~19 GB at 4-bit |
| License | Apache 2.0 | Apache 2.0 |
| Runs on | 24 GB GPU | 24 GB GPU |
| Our rank | #5 | #6 |
Shaded cells win the row, with prices read on November 18, 2026.
When to choose Qwen 3.8 27B
Choose Qwen 3.8 27B for the most room left for context on one 24 GB card.
When to choose Gemma 4 31B
Choose Gemma 4 31B Heretic if you prefer Google's Gemma base, or a build with settings searched by Heretic.
Run them
Load both as GGUF builds in Ollama, LM Studio or llama.cpp. When a job outgrows one GPU, run DeepSeek V4.1 Flash Abliterated by API.