AI Wiki — Knowledge Base

Can a relatively small language model (Qwen 3.8 27B) understand visual style and not just describe it?I asked it to recreate 618 visual language of designers, artists, movements etc and then transfer those ideas into entirely new SVG artworks and pages. Here what we have:

I decided to do this experiment to see what would be the ability of an LLM model to, in a single shot, transfer the knowledge it may or may not have about that artist/brand/designer/animator to a comp

model-releasesr-localllama

21 Sept 2026

Search: “r-localllama”

Nathan Lambert's written Congressional testimony on the state of open models - Chinese open-weight downloads now 2x America's, >80% of OpenRouter open-model usage

New user - seeking help

open-webui and open-terminal thoughts

Perhaps the highest quality mainline quants of Qwen3.8 27B?

Pi agent qwen 3.8 flash next plays Baldur's Gate 2

Qwen 3.8 Flash Next q4_k_m, 130k context, q8 cache on 16GB VRAM ann 64GB RAM, 15-20 t/s on 4080

Qwen FN vs 27B --- Think I'm saturated.

Streaming Nemotron 3 Diarization

stuntd: a local Jev-compatible server on Laya that learns from your own traffic (no API key needed)

The attack on open weights continues to grow.

Am I going insane for thinking that these are all Bot comments?

Cost of a local set-up for coding?

Dynamic Quantiser - a way to make your own high quality dynamic quants

Fork of FreeToken with DeepSeek-V4.1, vision and speculative decoding (2x3090 numbers inside)

Is llama.cpp meant to be slow at long context, even when you aren't using that context?

Kev: tiny Jev-like decision models (0.8B/4B/9B) on Qwen3.5 you can train and run locally - the 9B fits a 32GB Mac

MiMo-V2.6-Flash on vLLM: fixes for 'empty responses' with thinking + tools, and a hidden 2,048-token output cap

mini-AGI: Continual-learning dynamically looped transformer with evolutionary grown (on a laptop)

Model grafting: turning Qwen3.5-4B into a causal encoder-decoder after the fact

New 6B image model coming, AntLing just open sourced the Ming-Image-0.1-Design family

Ngram and world knowledge - why are we just building a coding model?

quants for K2-Horizon are now available

Qwen 3.8 27B at ~3 BPW on an RTX 3060: GSQ vs ByteShape IQ3-XXS 2.88BPW

Qwen3.8-27B: >70 tok/s (>160 tok/s concurrent), 10k tok/s prefill, full context on 2x3090 (or and 48GB or larger on ampere or higher), vanilla vllm

[R] Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention

What are normie reactions to your local AI use like?

What underrated AI tools have actually made you more productive in 2026?

XiaomiMiMo/MiMo-V2.6-Distill-Qwen-9B

16GB (and in many cases 12GB) is the max vram most people will ever reasonably have

2x r9700 in one strix halo machine

50+ Hours and 100M+ Tokens Later, Open Source Autonomous Agent is GETTING CLOSER at Solving an Open Math problem

A better coder for the small-GPU/small-RAM crowd!

Can a relatively small language model (Qwen 3.8 27B) understand visual style and not just describe it?I asked it to recreate 618 visual language of designers, artists, movements etc and then transfer those ideas into entirely new SVG artworks and pages. Here what we have:

convaiinnovations/laya (multilingual, non-autoregressive System 1 decision model)

Deepseek training 2T and plans 8T model

Deterministic Kittens: fun with Qwen Image 2.1 on an M2 Macbook Pro

Gewell - Gemma4 inference engine

Got jev-like api running natively on ninfer / qwen3.8 27b and results are quite decent

Has anyone tried ingesting 1k-100k scientific papers into LLM wiki?

hearim(헤아림): Maybe you don’t need a special model for Jev — ordinary local LLMs already have the capability

Hemmingway-1: An AI that writes like a person (Qwen3.8-27B finetune)

[MASSIVE RELEASE] Supra2-IMG - a tiny 100M text-to-image model - SOTA quality and open release!

Mimo v2.6-Flash-RL vs open-weight models

mini-AGI - dynamically grown (530M params currently and growing) continual learning model trained from scratch on 8GB VRAM laptop from batch-1 stream of data.

Putting the question before the context took my local Qwen from 89% to 100% on a decision benchmark, and from ~400 ms to ~80 ms

[Splash Engine] Qwen3.8-27B in native 8-bit at 37–55 tok/s on Apple Silicon: Extending Splash to Q8, 256k context scaling, and the 'Reasoning Cliff'

Uncensor an LLM without touching weights: inject a tiny trained KV-cache bank (~18MB) and unload it anytime

Where are the current GPU VRAM sweet spots?