AI Wiki — Knowledge Base

R9V Update: Created and adopted KVA projections based on Deepseek V4.1 Flash + HySparse2/MiMo-V3 for Qwen3.8 Flash Next. This is a game changer for models that don't natively implement it. 1.45-1.85x speedup in prefill to 3k+ at a small deficit to perplexity. [2x R9700, 128GB DDR5]

Here's my *first* implementation of KVA projectors on QFN (just the uncensored model for now) the highlights are basically as follows for using the projectors at each different layer: Starting at laye

model-releasesr-localllama

24 Sept 2026

Search: “models”

Clock Diffusion: Efficient Semi-Autoregressive Continuous Diffusion Language Models

When Do Biological Reasoning Models Use Their Biological Inputs?

World Motion Models: Flexible Sequence Modeling of SE(3) Trajectories

Searching for BSM Experimental Signatures with Large Lagrangian Models

A Model-Agnostic Physics-Guided Adapter for Few-Shot Transfer of Coastal Flood Prediction Models to Unseen Regions

How Language Models Differ in Redistributing Attention-Head Activity Under Serial Demand

Gentle reminder that Hugging Face was declined help from Anthropic & OpenAI models when they were being hacked by OpenAI models Self-hosted …

Auditing System-1 Models on Biosecurity-Relevant Benchmarks: Calibration, Selective Prediction, and Permutation Instability in a Non-Generative Model

PALM: Point-in-Time Adaptation for Financial Language Models

R9V Update: Created and adopted KVA projections based on Deepseek V4.1 Flash + HySparse2/MiMo-V3 for Qwen3.8 Flash Next. This is a game changer for models that don't natively implement it. 1.45-1.85x speedup in prefill to 3k+ at a small deficit to perplexity. [2x R9700, 128GB DDR5]

On the Effect of Bit-Level Parameter Perturbations in Machine Learning and Deep Learning Models

On the Limitations of Large Language Models for Conceptual Database Modeling

BOOM: Benchmarking Out-Of-distribution Molecular Property Predictions of Machine Learning Models

Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models

Drive-HWM: Hierarchical World Models for Dynamic-Latent Guided Autonomous Driving

FPCO-Dialog: A Multi-Turn False-Premise Benchmark for Correction and Cooperation in Vision-Language Models

Performance Foundations of Parallel & Distributed Reasoning Language Models

VIPER: An Expert-Curated Benchmark for Vision-Language Models in Veterinary Pathology

No, Engrams won't let you run 1T models locally. It does something even better.

Design-to-Plan: A Large Language Model-Based Multi-Agent Framework for Manufacturing Process Planning from 3D CAD Models and 2D Engineering Drawings

A look at startups like General Intuition working on large action models, aka world models, which are trained on videogames and simulations, to pilot robots (Christopher Mims/Wall Street Journal)

FlavourBench: Ranking Frontier Language Models with Executable Culinary Ground Truth

[Paper] ToMoE: Converting Dense Large Language Models to Mixture-of-Experts through Dynamic Structural Pruning

LiveHouse-TS: An Open-world Living Benchmark for Time Series Foundation Models

LiD-GLM: Lipschitz-constrained Deep Generalized Linear Models

Attributing Preprocessing Invariance in Spectral Foundation Models

Small models. Big cost savings. @relace_ai trains specialized models purpose-built for code generation, powered by the dedicated @ycombinato…

Continuous-Latent Predictive Modeling with Semantic Alignment for EEG-Language Foundation Models

TACTICL: Task-Aware Compression of Tabular ICL Models

For a very long time most high-performing AI models were end-to-end neural models; vector input -> vector output, with only ultra-thin symbo…

Developers in Africa are increasingly choosing Chinese open-source AI models over US models, saying they are downloadable, easier to customize, and much cheaper (New York Times)

Revisiting Predictive Process Monitoring in the Age of Foundation Models: A Comparative Study of Sequence, Tabular, and LLM Approaches

With release of Deepseek V4 I wanted see how the model sizes are trending over time. The trend is that by this time next year, we probably will have Opus 4.5 level models on consumer grade laptops!

Tag Questions and the Generational Reversal of Sycophancy Across 45 Language Models

Robot-Factored World Models via Robot Rendering

Small Vision-Language Models Know When They Are Wrong But Cannot Say So: A Two-Model Study of Stated versus Internal Confidence Under Realistic Image Degradation

100B Models on Cheap Hardware: how realistic and limitations

Yes, open-source / open-weight models are important for a healthy AI ecosystem. That's how we can verify things, check claims, and keep up o…

[Paper] Statistically-Lossless Quantization of Large Language Models

Interesting finding on frontier models. It turns out that frontier models can write proofs but stumble on faithfully copying a long block of…

Collective Intelligence with Foundation Models

Grok 4.5 and Composer are two different model weight classes. Composer 2.5 will remain offered, and we will release new models of this size …

GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation

Optimal Mixture-of-Experts Model Averaging for Conditional Generative Models

Tripo AI, which is developing 3D foundation models and world models for use in games, raised 150M, a month after it said it raised 200M (Dean Takahashi/GamesBeat)

[Thread] The US' now-lifted export controls on Anthropic models created uncertainty, are an own goal, and will push service providers towards Chinese AI models (Alex Stamos/@alexstamos)

When Models Know When They Do Not Know: Calibration, Cascading, and Cleaning

Thoughts: 1. Harness > models 2. Enterprises will jump on using zenith-shaped things to optimize costs (esp. with open source models) Can't …