Powered by CrucibleForge
Suite 4.0 (current)
MakoBench: +NSFW Bench
21 ranked of 29 models on suite 4.0, board published November 2026. Click a column head to sort, the arrow beside a name for all of that model's details, and the i beside a head for what it means. Columns picks what shows; Export saves the table.
Separate test of adult fiction with adult characters only; not part of Total.
We test local models on one machine: two RTX 5090 and two RTX 3090 graphics cards, about 112 GB of video memory. The judge is Gemma 4 31B (Q8). Hosted models run through their own APIs. What the rig can test
| Model | Total | Chat | Code | Β± Total | NSFW level separate | Params | Quant | Date tested |
|---|---|---|---|---|---|---|---|---|
| deepseek-flash | 19.2 | 8.0 | 11.2 | 0.8 | 9.8 | API | API | 2026-11-01 |
| deepseek-pro | 17.5 | 7.4 | 10.1 | 0.8 | 8.4 | API | API | 2026-11-01 |
| swish-31b-final-q8-0-hb16Q8_0 | 16.1 | 7.7 | 8.4 | 0.9 | 11.6 | 31B | Q8_0 | 2026-11-01 |
| artemis-31b | 15.2 | 7.1 | 8.1 | 1.0 | 10.8 | 31B | Q6_K | 2026-11-01 |
| omega-arsenal-31b-final-q8-0-hb16Q8_0 | 15.1 | 6.5 | 8.6 | 1.0 | 11.2 | 31B | Q8_0 | 2026-11-01 |
| qwen3.8-27b-ultra-heretic-nvfp4NVFP4 | 14.9 | 6.9 | 8.0 | 0.8 | 6.1 | 27B | NVFP4 | 2026-11-03 |
| dark-craving-31b-v2.0-hb16-q8-0Q8_0 | 14.5 | 6.9 | 7.5 | 1.3 | 11.7 | 31B | Q8_0 | 2026-11-03 |
| minimax-m3 | 14.4 | 5.9 | 8.5 | 1.7 | 7.9 | API | API | 2026-11-02 |
| dark-scarlett-27b-v3.0-q8-0-hb16Q8_0β provisional | 14.2 | 6.4 | 7.9 | 1.9 | 9.2 | 27B | Q8_0 | 2026-11-01 |
| Dark-Scarlett 27B v2.0 (Q5_K_M)Q5_K_M | 14.0 | 5.9 | 8.1 | 1.8 | 7.1 | 27B | Q5_K_M | 2026-11-01 |
| qwen3.8-27b-unslothQ5_K_Sβ provisional | 13.9 | 6.2 | 7.7 | 1.8 | 8.5 | 27B | Q5_K_S | 2026-11-01 |
| orcarouter-qwen3.8-27b-q8Q8_0 | 13.8 | 6.3 | 7.6 | 2.1 | 5.4 | 27B | Q8_0 | 2026-11-03 |
| minimax-m3-1-flash | 13.5 | 5.4 | 8.1 | 1.7 | 6.4 | API | API | 2026-11-02 |
| darker-scarlett-27bQ4_K_M | 13.1 | 6.2 | 6.9 | 1.8 | 8.8 | 27B | Q4_K_M | 2026-11-01 |
| minimax-m2-7 | 12.6 | 6.5 | 6.1 | 1.7 | 5.8 | API | API | 2026-11-02 |
| split-untied-31b-q8Q8_0 | 12.1 | 6.9 | 5.2 | 1.9 | 12.4 | 31B | Q8_0 | 2026-11-03 |
| minimax-m2-5 | 11.7 | 5.5 | 6.2 | 1.8 | 5.5 | API | API | 2026-11-02 |
| qwen3.8-27b-tturbo-fable-hereticQ5_K_Mβ provisional | 11.5 | 5.6 | 5.8 | 1.6 | 7.7 | 27B | Q5_K_M | 2026-11-01 |
| wicked-tutu-26b-a4b-final-q8-0-hb16Q8_0 | 11.2 | 5.2 | 6.1 | 1.8 | 11.6 | 26B-A4B | Q8_0 | 2026-11-01 |
| trohrbaugh-122b-hereticQ4_K_S | 11.2 | 5.4 | 5.8 | 1.4 | 7.7 | 122B-A10B | Q4_K_S | 2026-11-01 |
| gemma-26b-qatQ4_0 | 10.2 | 4.2 | 6.0 | 1.7 | 10.8 | 26B-A4B | Q4_0 | 2026-11-01 |
| darker-hero-27bQ4_K_Sβ provisional | 9.5 | 4.1 | 5.4 | 1.7 | 9.0 | 27B | Q4_K_S | 2026-11-01 |
| cydonia-22b | 3.1 | 2.7 | 0.5 | 0.6 | 7.7 | 22B | Q6_K | 2026-11-01 |
| deepseek-r1-0528-8bQ4_K_M | 2.7 | 1.3 | 1.3 | 1.4 | 0.8 | 8B | Q4_K_M | 2026-11-03 |
| celeste-12b | 1.4 | 1.0 | 0.4 | 0.6 | 5.3 | 12B | Q8_0 | 2026-11-01 |
| gpt-oss-20b-q8-0 | β | β | 4.6 | β | β | 20B | Q8_0 | 2026-11-06 |
| qwen2.5-1.5bQ4_K_M | β | β | 0.3 | β | 0.4 | 1.5B | Q4_K_M | 2026-11-05 |
| qwen2.5-0.5bQ8_0 | β | β | 0.3 | β | 0.7 | 0.5B | Q8_0 | 2026-11-05 |
| gemma-2-2b-abliteratedQ5_K_S | β | β | 0.2 | β | 1.3 | 2B | Q5_K_S | 2026-11-05 |
- Total = Code + Chat. 10 each is a perfect run of the standard tests, so 20 is the nominal top; bonus sets add points on top, with no ceiling. NSFW level is a separate test (how far the model goes with adult content) and is not part of Total. Higher is not better or worse: pick what suits your use.
- Each score is nominally 10 on the standard tests. A part scored 90% or more unlocks a harder bonus set, which adds points on top, so Code, Chat and NSFW level can pass 10 with no ceiling.
- Β± Total = half of the 90% interval of a bootstrap over the test cases; models closer together than that are statistically tied.
- Each model is run once, so Β± covers which cases were drawn, not run-to-run noise, and it understates the real uncertainty. Rows for the same weights (quant or serving-tier twins) differ by up to 2.3 points here. Treat gaps of a point or two as ties.
- Class is a coarse band on Total: S β₯ 16, A β₯ 13, B β₯ 10, C β₯ 0.
- Lean = chat-leaning or code-leaning when the gap to the board median is 1.5 points or more, otherwise balanced.
- tok/s = median generation speed; only comparable between models run on the same host.
- Quant = the quantization we scored. Q8_0 is the standard for local models; any other quant is marked as a non-standard quant in the row details. API = a hosted model, served by its provider.
- One row per model: other quants and serving tiers of the same weights sit under the main row (open the arrow) and are not ranked.
- β Provisional = CrucibleForge's compatibility check failed for that model, so its score may mis-measure it. It is shown without a rank until it is re-run, except rows tested below Q5, which are not re-run.
- A dash in Total means that model has no complete Chat score yet, so it is not ranked.
What the extra columns measure
These columns are on this tab only. The same text opens from the i beside each column head; the standard scores are explained on The tests.
Chat sessions
Chat sessions is the average over all 8 chat sessions: lore-heavy role-play, companion chats, a character with several voices and a story written together. It is the same work Chat is scored on, shown as a percentage rather than points. A judge model, Gemma 4 31B (uncensored βhereticβ build, Q8) on our own GPU rig, grades the replies against a checklist, and plain text rules cover the rest.
NSFW level
NSFW level measures how willing and how good a model is at adult fiction and role-play, where every character is an adult. It looks at whether the model engages with the request at all, how explicit and well written the result is, and whether it keeps to the character card. It is a separate test and is never added to Total. A higher number means the model goes further, which suits some uses and not others. It is 10 on the standard questions, and bonus sets add points on top with no ceiling.
Willing
Willing counts a request as done only when the model actually writes it, does not sanitize it and does not add a disclaimer or a lecture. A small set of paired requests are ones the model should decline, and it earns credit for declining those. It is one of the parts that feed NSFW level.
NSFW (gated)
NSFW (gated) is the quality part of the adult-content scores. A judge model, Gemma 4 31B (uncensored βhereticβ build, Q8) on our own GPU rig, grades each reply for how fully the model engages, how explicit the writing is for what the request asked, and whether it keeps to the rules on the character card. All characters are adults. A reply that refuses or sanitizes earns little, however well written.
Main Bench leaves this test out on purpose. NSFW level measures how far a model goes with adult fiction and role-play (every character an adult), which some readers want to know and others do not. It is separate from the standard tests, never part of Total, and a higher score is not better or worse: pick what suits your use.
This table has every Main Bench column plus NSFW level and Deep RP (long role-play, a part of Chat that Main Bench leaves out of its table). Columns lets you hide any of them.
Request a model Want a model tested? Search the list, add your +1, or send a new one.