Is there an uncensored MiniMax-M3?
MiniMax-M3 from MiniMax is ranked 13 of 100 distinct active open-weight models in this Artificial Analysis snapshot.
Research and availability status
This is a research guide, not an available Wild West API model listing. No uncensored or abliterated MiniMax-M3 endpoint is confirmed here. An open-weight entry on AA does not establish that a derivative exists, works with a particular harness, or is hosted by Wild West API.
“Uncensored” describes reduced refusal behaviour or a service’s policy, not a standardized editing method. “Abliterated” refers to an intervention intended to reduce refusal-related behaviour through model-internal edits.
What rank 13 actually tells you
MiniMax-M3 scores above 87 of the 100 selected entries: a strictly-lower-score percentile of 87% in this snapshot. No other selected entry shares its exact index score. This entry is not flagged as an estimated index in the source data.
It sits in the upper half of this selection. The relevant decision is whether its measured strengths fit the workload better than adjacent entries, not whether its rank is the highest available. Hosting cost and hardware requirements are outside this snapshot.
Base-model benchmark evidence
These are measurements attached to the evaluated base model, not an uncensored derivative. Missing values mean not reported here, never a zero score. Same-benchmark comparisons below use only entries with a reported measurement.
| Benchmark | Reported result | Within-snapshot comparison | Task coverage |
|---|---|---|---|
| GPQA | 92.9% | 93 of 96 measured entries score lower on this same benchmark | graduate-level science questions |
| Humanity’s Last Exam | 39.0% | 92 of 100 measured entries score lower on this same benchmark | a broad set of difficult expert questions |
| SciCode | 47.1% | 42 of 56 measured entries score lower on this same benchmark | scientific coding tasks |
| IFBench | 82.9% | 64 of 65 measured entries score lower on this same benchmark | instruction-following constraints |
| τ²-Bench | 88.9% | 57 of 64 measured entries score lower on this same benchmark | interactive agent tasks |
| Terminal-Bench Hard | 42.4% | 61 of 63 measured entries score lower on this same benchmark | hard terminal-based tasks |
The highest numerical reported percentage for MiniMax-M3 is GPQA at 92.9%; the lowest is Humanity’s Last Exam at 39.0%. These are numerical extremes, not evidence that one skill is stronger than another: benchmark percentages cannot be compared as if the tests had equal difficulty, scoring rules or task coverage. Compare GPQA with that same benchmark on another model, not with a different test.
Context window and a practical evaluation workload
At 1,000,000 tokens, a candidate workload is a large document collection or a repository with supporting specifications. A useful test would ask the model to trace a claim across distant files, then verify every citation. A large advertised window does not establish accurate retrieval throughout that window, and packing it fully can change latency and cost.
81 of the 100 entries have a smaller reported window than MiniMax-M3. As a planning example, allocating 75% of its reported window to source material gives 750,000 tokens, leaving 250,000 for instructions, conversation, tool messages and output combined. This is arithmetic for a test budget, not an endpoint limit or an assertion that output can use the entire remainder.
Reasoning settings and release comparisons
It is marked reasoning by AA. There is no within-model setting spread to quantify, so a claim that a lower budget preserves its score would be unsupported. A derivative should be tested in the actual setting you intend to use.
No other selected entry is credited to MiniMax in this 100-model snapshot. There is no publisher-peer comparison to make here, and similarly named releases outside this selection should not inherit these measurements.
Nearby ranked alternatives
The closest ranks give a bounded shortlist for evaluating MiniMax-M3. A small index gap is not a statistical significance claim; compare task outcomes and deployment conditions directly.
| Alternative | Rank | Index and gap from this model | Context |
|---|---|---|---|
| Motif 3 | 11 | 33.57 (+4.35 points) | 262,144 tokens |
| K2 Horizon 375B A23B | 12 | 30.50 (+1.28 points) | 524,288 tokens |
| Nex-N2-Pro (Based on Qwen3.5-397B-A17B) | 14 | 28.20 (-1.02 points) | 262,000 tokens |
| Kimi K2.7 Code | 15 | 25.81 (-3.41 points) | 256,000 tokens |
How to evaluate an uncensored or abliterated candidate
For MiniMax-M3, first locate the exact base release and any claimed derivative’s model card. Verify access, redistribution and usage terms at their original sources: AA’s open-weight classification does not assert an open-source license. Record the base revision, derivative revision, runtime and reasoning setting so your comparison can be reproduced.
Weight-based abliteration generally studies differences between activations on refusal-inducing and ordinary prompts, identifies candidate refusal-related directions, and intervenes on model internals. It is not equivalent to changing a system prompt or removing an API filter. Whether a particular intervention is supported depends on the model and implementation; this guide supplies no architecture-specific recipe or claim that the method has been validated on MiniMax-M3.
Run the base and candidate on the same permitted task set before and after changes. Measure refusal behaviour separately from factual errors, instruction following, coding correctness and tool execution. Use GPQA as one reference for task coverage, while keeping its published percentage separate from your own results. Track any regression and repeat with the exact context length and setting your application needs. A lower refusal rate is not evidence that all other behaviour survived unchanged.
Sources and scope
Snapshot retrieved 2026-10-06T06:17:07.812043+00:00. AA’s MiniMax-M3 model entry supplies the evaluated setting and measurements; the Artificial Analysis leaderboard is the population source. Top 100 distinct active open-weight model names by Artificial Analysis Intelligence Index. Reasoning-budget parentheticals are removed for grouping; the strongest scored setting represents each group. Open weights is AA's classification, not an open-source-license assertion. Scores describe the evaluated base models, not uncensored derivatives. Rank is within this selected active open-weight population, not the complete leaderboard.
For the research behind refusal-direction interventions, see Arditi et al., Refusal in Language Models Is Mediated by a Single Direction. This general research source does not establish successful abliteration of this release. Continue with the abliteration explanation or the current Wild West API line; only the latter lists products for sale.
Questions about MiniMax-M3
Is MiniMax-M3 abliterated on Wild West API?
Not confirmed by this guide. This page documents an AA open-weight research entry, not a hosted derivative. Check the current model line for available products.
Does the 29.22 index score apply to an uncensored derivative?
No. It describes AA’s evaluated MiniMax-M3 base-model setting. Changes to weights, prompts, quantization, tools or reasoning budgets require fresh evaluation before carrying the score over.
Can I use the full 1,000,000-token window?
That is the context figure in the AA snapshot, not a guarantee for every deployment. Confirm the actual endpoint or runtime limits and reserve space for output and tool messages. Test retrieval accuracy near the ends of long inputs.