It has been 4 years since c.ai launched the first model with hallucinations, yet we are still nowhere near where we should be when it comes to the liveliness and entertaining hallucination capabilities of large language models. Fortunately, new architectures that emerged last year now allow users with 10–16GB of VRAM to run genuinely usable models with context windows of up to 64k via Kobold or LLAMA.
Hardware Requirements: 10–16GB VRAM – 26–32GB RAM.
Evaluation Criteria:
- Speed: Whether generation speed hits at least 10 t/s.
- Coherence: Whether dialogues, monologues, and perspectives make internal sense beyond basic robotic logic or plain behaviorism.
- Japanese / Brazilian Portuguese / American English: Ability to handle and comprehend typos, subtle semantic nuances, or sharp phrasing.
- Slop Patterns: The frequency of cliches ranging from “casting long shadows” to “smells like ozone and burnt sugar.”
- MoE vs FULL: Preference for Mixture-of-Experts (MoE) over full dense models.
- Censorship / Refusals: All of these models have essentially zero or negligible refusal rates, even when dealing with NSFL topics, so there is nothing to worry about.
- Instruction / Prompt Engineering: Tested both with and without the Pandora instruction set; if a model can handle 23 interwoven rules, it can tackle practically any task without even needing prompt engineering.
Side note: I skipped all models with Chain-of-Thought (CoT)—it’s a pure waste of tokens, noticeably slower, and completely unnecessary. A well-crafted prompt yields the exact same results. Just pick quants higher than Q_4_K; size doesn’t matter ( ͡° ͜ʖ ͡°).
The Catch: I ran a quick trivia benchmark across 70B, 120B, and 300B models on anime, manga, movies, and music. Turns out only 200B+ models actually have this kind of knowledge baked in—even 120B models barely contain the pop-culture lore you’d care about. But then again, who can realistically run those?
Model List (Ranked Best to Worst / Reasoning):
- Gemma 4 26B-A4B: Comparable to Qwen 35B-A3B, but delivers better dialogue/monologues and instruction following; slightly smarter, sharper, and more engaging.
- Qwen 3.6 35B-A3B (Abliterated): Fast, accurate, and quite fun. However, without solid guiding instructions, it can overfit to previous input/output formats and patterns, or ramble and ask too many questions due to instruction issues.
- Gemma 4 Instruct 19B-A1B Heretic (Abliterated): On par with the 26B-A4B, just slightly more rigid and flat, though it maintains strong accuracy and logic.
- The Omega Directive 14B: Follows instructions well, low slop rate, excellent prose, but its generations can become formulaic once the context fills up a bit.
- GLM-4.7-Flash 31B-A3B (Abliterated): Performs similarly to Qwen 35B-A3B, but more blunt and slightly worse at grasping subtle semantic nuances and phrasing.
- Qwen3 30B A1.5B 64K High Speed NEO MAX: Basically the 35-A3B with enhanced reasoning capabilities.
- MN CaptainErisNebula Chimera Heretic Uncensored 12B (Abliterated): Rather bizarre behavior—a bit eager to dump everything all at once and loaded with slop. That said, it’s the most proactive of the bunch, rarely gets stuck in loops, and has solid dialogue/monologues, though it’s ill-suited for Q&A-style characters.
- Mistral Nemo Instruct 2407 HERETIC HI Claude Opus 12B: Decent performer, with cliches reasonably polished. However, it tends to overfit to the RP perspective or constantly try to flirt with you; not ideal for Q&A characters.
- Gemma 3 12b GLM-4.7 Flash INSTRUCT Hybrid Heretic (Uncensored): Fun RP persona and accurate, but prone to overfitting to previous input formatting/style, frequent slop, or being overly humorless/uptight.
- Ministral Instruct 2512 Absolute Heresy 14B: Mistral is terrible at adhering to complex rules (regex, status bars, directives, lorebooks, etc.) or constantly turns {{char}} into a pickup artist. Ironically, it’s the absolute best in terms of dialogue, monologues, and sheer personality.
- GLM 4.6V Flash Hybrid ?B: Similar to the UD version, but with less slop and even fewer refusals.
- GLM 4.6V Flash UD 9B: Depending on how you define {{char}}‘s personality, it can feel robotic/repetitive and tends to ask too many questions, despite stellar accuracy in following all 23 PI rules + regex instructions.
- Gemma 4 E4B UD: A budget compromise—super lightweight, fast, accurate, and smart, but prone to looping/slop patterns.
- LFM2 12B A1B SpeedDemon The Deckard II HERETIC (Uncensored): Blazing fast, accurate, and smart, but lacks flavor and soul.
Overall Impressions: GLM leans pessimistic and overly analytical; Qwen tends to be overly curious and loves purple prose; Gemma is usually witty and creative; Mistral leans flirtatious with bizarrely specific details; Llama is typically slop-heavy and predictable. This holds true across virtually all of their fine-tunes, whether they are Omega, Heretic, Deckard, Dark, Horror, RP, Style, etc.
Running Setup: Kobold cannot load these models out of the box; use LLAMA for QWEN 35B (20.6GB) and GLM 31B (17.7GB). All other models were run at max speed with 40–64k context—with FlashAttention / FastForwarding / SWA / SmartCache + MMPROJ + TTS (if applicable) enabled in Kobold.
Known but discarded after testing (either too slow/large, poor performance, or had better alternatives): Mistral 24B, GPT 20B OSS, Kimi K2, Phi-4, Skyfall, DeepSeek, Cydonia, Magistry, Magnus Cydons, etc.
Untested Models: Goetia 26B-A4B v1.7, Gemma 4 26B-A4B StyleTune V2, Promethean Dawn 26B-A4B, and Fenrir-X 26B-A4B.
What if I have 24GB VRAM (e.g., RTX 3090)? If you want slightly smaller models, you can try: Omega Darker Gaslight The Final Forgotten Fever Dream 24B, Qwen3 24B-A4B Freedom HQ Thinking Heretic NeoMAX-D_AU, Cydonia 24B, Magistry 24B, Magnus Cydons 24B, Erebus 12B Instruct 2608, Mistral RP 24B, DeepSeek 33B, Velvet Eclipse 2x12B, or Velvet Eclipse 4x12B.
End
Top Discussions
ttkciar (⬆37): I spotted a few mistakes in this post (hallucinations, maybe?). For instance, I can’t find Gemma-4-19B-A1B-it anywhere. The closest thing I found on Hugging Face is https://huggingface.co/potto007/gemma-4-19B-A4B-text-REAP-GGUF . If a 19B-A1B model actually exists, I’d love to test it out. OP, could you double-check or update the post?
Manerfish (⬆23): Why are there ancient models and weird fine-tunes in here, including one by DavidAU of all people, lmao.
Atretador (⬆18): This feels like asking ChatGPT on the web to recommend runnable models, and it starts throwing year-old relics at you.
paq85 (⬆136): 1. Qwen 3.8 27b. 2. No point wasting your time 😅
tableball35 (⬆58): Though this post seems geared towards RP, and Qwen 27B isn’t really great at that, which is probably why it was excluded.
Pyros-SD-Models (⬆14): Qwen 3.8 27B has tons of insane god-tier fine-tunes for RP—it’s seriously mind-blowing.
dampflokfreund (⬆17): Running that requires 24GB VRAM, otherwise you’re crawling at 1–5 tokens/s. Hard pass.
Witty_Mycologist_995 (⬆13): If you’re looking for creative writing/RP fine-tunes, check out the Drummer series models or StyleTune.