Gemma 4 is Google's open weight family: a dense 31B model and a 26B mixture of experts with about 4B active parameters, both small enough to run on one GPU. Google ships them with strong safety tuning, so the community has produced abliterated, heretic and fine tuned versions that answer what the originals refuse.
The versions differ in how the refusals were removed. Heretic versions run an automated abliteration that searches for the edit with the fewest refusals and the least drift from the original. Others are fine tuned on compliant data, and some, like the RP versions, are tuned for roleplay on top of that. Those choices show up in agent reliability and in how much each version still refuses, which is what the table below measures. Why the method matters is covered in do abliterated models lose intelligence.
You do not need to host them yourself. The versions below run on the OpenUncensored API with an OpenAI compatible endpoint, and we measure them on the same tests as every other model. Every Gemma model we serve, ranked or not, is on the models page. Versions marked "not ranked yet" have refusal results but no agent run so far.
The originals
- Gemma 4 31B scores 24% on Humanity's Last Exam and 86% on GPQA Diamond (Artificial Analysis, reasoning mode).
- Gemma 4 26B A4B non-reasoning scores 12% on Humanity's Last Exam and 71% on GPQA Diamond (Artificial Analysis, reasoning mode).
- Gemma 4 26B A4B scores 19% on Humanity's Last Exam and 79% on GPQA Diamond (Artificial Analysis, reasoning mode).
Every uncensored version keeps that starting point. What differs is how much each edit costs in agent reliability, how much it still refuses, and the price.
Every uncensored Gemma 4 on OpenUncensored
| Model | id | Rank (score) | Agent tasks | Loops | Answers edgy requests | Hard-line refusals | $ per 1M in + out | Context | Providers |
|---|---|---|---|---|---|---|---|---|---|
| Gemma 4 31B Gembrain Uncensored Heretic | gemma-4-31b-gembrain-uncensored-heretic |
#14 (58.5) | 3/5 | 40% | 89% | 0% | $0.612 | 262K | nanogpt |
| Gemma 4 Uncensored | gemma-4-uncensored |
#19 (54.6) | 2/5 | 20% | 100% | 0% | $0.662 | 256K | venice |
| Gemma 4 26B A4B Uncensored TEE | gemma-4-26b-a4b-uncensored |
#20 (52.8) | 1/5 | 40% | 100% | 0% | $0.850 | 262K | nanogpt, redpill |
| Gemma 4 26B A4B Chimera X Heretic | gemma-4-26b-a4b-it-chimeraxheretic |
not ranked yet | – | – | 100% | 0% | $0.500 | 262K | nanogpt |
| Gemma 4 31B SDFT Heretic RP | gemma-4-31b-sdft-heretic-rp |
not ranked yet | – | – | 100% | 17% | $0.612 | 262K | nanogpt |
| Gemma 4 31B Cognitive Unshackled | gemma-4-31b-cognitive-unshackled |
not ranked yet | – | – | 100% | 100% | $0.612 | 262K | nanogpt |
Which one to use
- Best overall: Gemma 4 31B Gembrain Uncensored Heretic (rank #14, score 58.5).
- For agents and coding tools: Gemma 4 31B Gembrain Uncensored Heretic again, with 3 of 5 agent tasks finished and 40% loops.
- Most open: Gemma 4 Uncensored, answers 100% of edgy requests and refuses 0% of hard-line ones.
- Cheapest: Gemma 4 26B A4B Chimera X Heretic at $0.500 per million input plus output tokens.
Agent tasks are five multi-step tool scenarios, one run each, so a one task gap is noise. Answers edgy requests counts 36 legitimate prompts that cautious models refuse. The live numbers, including speed on real traffic, are on the rankings page.
Call it
curl https://api.openuncensored.com/v1/chat/completions \
-H "Authorization: Bearer $OU_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "gemma-4-31b-gembrain-uncensored-heretic", "messages": [{"role": "user", "content": "Hello"}]}'
If the providers for a version are down, the API falls back to the next best ranked uncensored Gemma 4 built on
the same original, and the X-OU-Model header tells you which one answered.
Try any of them in a minute
No account and no sign up. Buy a prepaid key at openuncensored.com/credits, or hand your agent openuncensored.com/agents.md and let it get one for you. Then:
curl https://api.openuncensored.com/v1/chat/completions \
-H "Authorization: Bearer $OU_API_KEY" -H "Content-Type: application/json" \
-d '{"model": "deepseek-v4.1-flash-1017-uncensored", "messages": [{"role": "user", "content": "Hello"}]}'
The full, always current table is on the rankings page.