Every Model That Can Be Run On 10-16GB VRAM Ranked
retained text
It's been 4 years since [c.ai](http://c.ai) first hallucination model, and yet, we're nowhere good enough at LLMs in terms of spontaneity/interesting hallucination features. Fortunately, last year brought new architecture that allow those within 10-16GB VRAM range to run something usable up to 64k Context Windows via Kobold or LLAMA.
**Requirements:** 10-16GB VRAM - 26-32GB RAM.
**How I Evaluated:**
* Speed; *If tokens per second are at least 10t/s.*
* Coherency: *If dialogue/monologue/POV makes any internal sense asides from basic logic (robotic feeling) or behaviorism.*
* JP/PT-BR/US-EN: *If it can handle/interpret typos, nuance or harsh verbiage.*
* Slop Patterns: *From "casting long shadows" to "smells like ozone and burnt sugar" frequency.*
* MOE vs FULL: *If rather pick a MOE instead of a FULL model.*
* Censorship; *All these models have either none or too low refusal rates even at NSFL topics, worry not.*
* Instruction/Engineering: *I've tested WITH and WITHOUT Pandora Instructions, if a model can handle 23 intertwined rules it can handle anything/not need a prompt at all.*
**Side Note:** I've ignored all CoT models, waste of tokens, slower, not needed, not at all. Do prompt nicely and it's the same. Just pick any quant higher than Q\_4\_K, size doesn't matter ( ͡° ͜ʖ ͡°).
**The Catch:** I've done a small knowledge test involving animes/manga/movies/music on 70B, 120B and 300B models and found out only models above 200B actually have this sort of data, even a 120B will rarely have the information about the media you're interest at. But who can even run it?
**The List (Best to Worse/Why):**
1. **Gemma 4 26B-A4B:** *Qwen 35B-A3B but with better dialogue/monologue, directive, slightly smarter/precise and funnier.*
2. **Qwen 3.6 35B-A3B (Abliterated):** *Fast, accurate, and decently fun, yet may overfit previous inputs/output format/patterns without proper directive instructions, or speak gibberish/ask too many questions because of directive.*
3. **Gemma 4 Instruct 19B-A1B Heretic (Abliterated):** *26B-A4B but a bit more rigid, plain, yet keeps accuracy and logic.*
4. **The Omega Directive 14B:** *Follows instructions, has low slop pattern rate and writes nicely, yet might become predicable after some context fill.*
5. **GLM-4.7-Flash 31B-A3B (Abliterated):** *Performs similar to Qwen 35B-A3B but more straightforward, less interpretative of nuance/verbiage.*
6. **Qwen3 30B A1.5B 64K High Speed NEO MAX:** *Basically 35-A3B with increased reasoning.*
7. **MN CaptainErisNebula Chimera Heretic Uncensored 12B (Abliterated):** *Its weird, a bit rushy trying to fill everything at once, full of patterns, but its the most proactive of them all, unlike to get stuck in a loop, has good dialogue/monologue yet ain't suitable for Q&A roles.*
8. **Mistral Nemo Instruct 2407 HERETIC HI Claude Opus 12B:** *It's decent, slop seems to have been rephrased properly, yet may overfit impersonation POV, or try to rizz you up mostly, not suitable for Q&A roles.*
9. **Gemma 3 12b GLM-4.7 Flash INSTRUCT Hybrid Heretic (Uncensored):** *Interestingly impersonative, accurate, yet may overfit previous inputs' formatting/style with frequent slop patterns or lack humour/too prim.*
10. **Ministral Instruct 2512 Absolute Heresy 14B:** *Mistral sucks at following complex rules (regex, status panel, commands, lorebook, etc), or turn {{char}} into a rizzler, despite being the best one regarding dialogue/monologue/spontaneity, ironically.*
11. **GLM 4.6V Flash Hybrid ?B:** *UD but with decreased slop frequency, even less refusals.*
12. **GLM 4.6V Flash UD 9B:** *Can feel robotic/repetitive depending on the personality of your {{char}} as making too many questions, despite high accuracy at following all PI 23 rules + regex commands.*
13. **Gemma 4 E4B UD:** *For the desperate, it's tiny, fast, accurate and smart, but will prolly repeat itself/have slop patterns.*
14. **LFM2 12B A1B SpeedDemon The Deckard II HERETIC (Uncensored):** *Lighting fast, highly accurate and smart, but lacks sauce.*
**Overall Review: GLM** tends to be pessimist and too logical, **Qwen** tends to be too inquisitive and fancy, **Gemma** tends to be funny and creative, **Mistral** tends to be flirty and oddly specific, **Llama** tends to be sloppy and predicable, this happens across all of their finetunes regardless if Omega, Heretic, Deckard, Dark, Horror, RP, Style, etc.
**Settings:** Kobold can't load these, use LLAMA for [QWEN 35B](https://www.paste.org/paste/view/0rbAY7U0) (20,6GB) and [GLM 31B](https://www.paste.org/paste/view/KdMODPPJ) (17,7GB). Other models all running at fastest speed possible with 40-64k - FlashAttention/FastFowarding/SWA/SmartCache + MMPROJ + TTS if any in Kobold.
Models I'm aware but discarded after trial cuz either too slow (massive), poor or have better alternatives: Mistral 24B, GPT 20B OSS, Kimi K2, Phi-4, Skyfall, DeepSeek, Cydonia, Magistry, Magnus Cydons, etc.
**Not Tested:** 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 going lower 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!<document fetched 2026-10-12 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1x3gpyi/every_model_that_can_be_run_on_1016gb_vram_ranked