My Little Awesome

Some AI Awesome Project list I've tried to use. Some my metrics attached, not all described (repo initially was for own purpose).

this cool **** was generated with❤️ by QwenIE 2511

My choose focus: Open source & preferably free or low cost, local-running, actual

Actual: 06.2025 (or maybe older) + little 05.01.2026 additions

All PR's and your foundations are VERY welcome!

General

LEADERBOARDS AWESOME

Agentic

Minecraft agents

(VLA or fast pipelines)

Good awesome summary here (OUTDATE = last updated 15.08.2025)

• CraftJarvis

  • 11.12.2025 глянуть OpenHA, тоже есть VLA, но медленные мб
    • TODO глянуть веса ЕСТЬ и код самое последнее вроде что у них есть
  • 31.07.2025 ROCKET-3 ждём веса ПОЛГОДА уже
  • 05.03.2025 2025 Jarvis VLA вышел, пока помойка (медленная и тупая), либо неправильно сделал ввод картинок через AutoClef

• 11.06.2025 Optimus 3 и мой форк

  • Works best for my scenario with AutoClef client minecraft agent, but I still don't know if it's random or prompt-following... It ignores instructions but looks like understands the game

• 2025 MrSteve

  TODO глянуть, но как будто бы используются базовые веса mineclip и steve1 VPT и особо нет связи с LLM

• Nvidia NitroGEN

  TODO глянуть, выглядит быстрой лёгкой и интересной! UPD: посмотрел, не впечатлил, слабый VLA, никакого коннекта с LLM, косячная либа, но универсален для многих игр и есть веса + маленький крутится на 4G VRAM. Можно адаптировать например под флеш-игры.

• DreamerV3

  • Will hard to implement, fully autonomous pipeline without direct goal conditioning, TODO test with client minecraft but looks unrealistic for now

• Game TARS (wait for releasing their weights!), should be good VLA instead of their computer-use slow agentic VLM

Computer-use agents

• ByteDance UI TARS Apache, SOTA?, very big and complex

  • Tested locally UI TARS 1.5 - fast & good, can simply play flash games and EVEN 100x slown-motioned minecraft
    • Launched on Agent TARS windows release 2.x versions
    • Has TERRIBLE repeating problem & ZERO reflection
    • still slow for realtime actioning

• Microsoft GUI Actor & Verifier & Critic old? Non-sota? Fara 7B? Fast but untested

• computer_use_ootb bad working, easy architecture, many platforms

OpenSource Local LLM RU

vsegpt ru - nice for few test runs

• Sainemo Remix I1 Q4 Overall good, better RU нашёл советы по параметрам: отсюда
• Saiga Nemo 12B Q4 Super entertain, mid quality, better RU
• Gemma 3 12B Q4 Overall good, normal RU, VISION
• Qwen 3 8B, 12B Q4 worse entertain, SOTA metrics, mid RU
• WAITING for weights of Gamayun RU LLM (paper), should be interesting
• FRED-T5-1.7B (FRED-T5-XL) Very GOOD creative multitask model but TOO OLD (T5 EncDec + GPT2 BPE tokenizer) + cannot convert to gguf
  • Best FineTune is SiberianFred

неплохие адаптации и ремиксы моделей под ру, сайга понравилась, остальные не знаю. + Устарел и не обновляется.

адаптация квенов под ру. В проде не запомнился, причин не помню, но занял место в старом лидерборде vikhrmodels

API LLM

Main famous - LMsys leaderboard

AA Leaderboard

Нашёл такой ещё лб пока не смотрел но там вроде есть что-то по бенчам

OpenSource Embeddings

[outdated]

• user bge m3 ru, tested, understanding, medium size
• nomic mostly en, tiny, ru poor quality

Agentic Workflow

[outdated]

• python + langchain
• n8n
• Google Opal?

Media

Multimodal discussion

[MAY BE OUTDATED]

• VideoXL SOTA for hour videos
• Gemma 3n only python image, video, audio, not precise
• InternVL3 Precize
• Qwen 2.5VL medium, partly precise

Face Swap

[MAY BE OUTDATED]

• FaseFusion slow SOTA FaceSwap, OpenRAIL license, Docker Use, Commerical added install, webcam use

• DeepLiveCam normal but need fix after install

• DeepFaceLive & DeepFaceLab Old archived VERY functional FaceSwap, not work for now? untested fork: MachineEditor/DeepFaceLab-MVE

VideoGen

• LTX-2 opensource

  semi-OS license, fast, SOUND-gen

  • Min 16G VRAM
  • buggy but prompt-following

• Wan 2.2 Commerical use allowed, VERY high quality, TOO weak hard prompt following (and sometimes simple...)

  • Best implementation - Phr00t/wan aio
  • Best on ~24G VRAM, can work on 16G and sometimes on 8G

ImageGen

imgsys ld, AA Leaderboard

• Perchance AI web, free, not censured, looks like old Stable Diffusion 1.5+lora. Chroma-based?
• Qwen Image Edit comfyui, 90s (50s with 4step fusion lora), edit, weak censure, good prompt follow PROD-USEABLE for few changes
  • QwenIE 2511 MUCH BETTER
• ZImage
  • Good Txt2Img, Bad controlnet, full uncensored
• Flux2 dev
  • TOO resource-intensive, good quality, a bit censored (in prompt-following)

TTS ru

• FishSpeech

  for now the best for RU, good accents, but non-commerical weights license

  • 1.5 good but fast on CUDA-only
  • s1-mini same but maybe worse...

• Silero

  very fast

  • Limited voicelist

Licenses understanding

Good understandable explanations https://choosealicense.com/licenses/

Commerical use allowed: Public domain, The Unlicense, Apache 2.0, MIT, ISC, GNU General Public License (Need same license): GNU LGPLv3, GNU AGPLv3, GNU GPLv3; Mozilla Public License 2.0, Boost Software License 1.0, CC (any without NC; ND & SA - same license-like restrictions) ...

Most of them need notices