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: “model-releases”

Live Assistant: Learning Whether, When, and Whom to Assist in Real-World Live Social Streams

Local Evidence and Geometric Readout Repair in Trained GNNs

Lost in Speech: Trilingual Spoken Hallucination Detection Across Audio and Transcripts

Mac Studio M5 Ultra 96GB vs M5 Max 128GB for local LLMs?

Math Reasoning in LLMs is Organized by Approach, Not Topic

Memory Attention

MessyKitchens: Contact-rich object-level 3D scene reconstruction

MINER: Multi-crop INference-time Enhancement for Rare-Object Retrieval with Frozen Dual Encoders

Minimal-Norm Univariate Two-Layer ReLU Classification: Exact Solutions and Global Optimality with Skip Connections

Mizar: A 159M-Parameter Audio-Language Model for Audio Understanding

MLPerf Training Introduces Its First LLM Post-Training Benchmark

model : add Ling 3.0 VL support by aetherbird · Pull Request #29151 · ggml-org/llama.cpp

MolDesignBench: Evaluating LLM-based Agent for Scenario-grounded Molecular Design

MSK-Bench: Benchmarking Full-Body Musculoskeletal Motor Control Across Tasks, Control Paradigms, and Physiological Metrics

MultiVENT-Raw: A Benchmark for Retrieval and Reasoning over Raw Videos

MWE-ECL: Recoverable Long-Range Context Does Not Always Override Local Lexical Priors

My foray into local ai. Two BC-250 ex mining apus running Qwen3.6-35B-A3B Q4_K_M at 60 tok/s with 64k context

NADI 2026: The Second Multidialectal Arabic Speech Processing Shared Task

Neither Silence nor Overlap Is Failure: Intent-Conditioned Evaluation of Turn-Taking in Full-Duplex Spoken Dialogue Models

Noob, Question.

NS-ATTENTION: Newton-Schulz Transformations of Attention Outputs in Vision Transformers

Omni-Decision: Evidence-Ledger Planning for Omni-Modal Agents

On Preference Coverage Collapse from Hindsight Relabeling in Multi-Objective Reinforcement Learning

Optimization without Future Compromises? Decentralized Coordination via Collective and Reinforcement Learning

Parameter Importance-Driven Continual Learning for Foundation Models

PASTABench: Proactive Assessment of Sequential Trajectories for Agent Safety

PBLH Estimation from Satellite Radiances via a Dual-Encoder Transformer

Phonemizing User-Generated Text: A Benchmark, Taxonomy, and Compositional Approach

PotARCin: Multi-Dimensional Evaluation of Skill Acquisition in Abstract Reasoning Tasks

Power your agents: Gemini 3.8 Live with Live Avatar is now generally available

PR-Smoother: Simulator-Preserving Non-Gaussian Smoothing for Data Assimilation

Predicting Activities in Aqueous Electrolyte Solutions with Hybrid Machine Learning

PRISM-VLM: A Multi-Axis Discriminative Benchmark for Compact Vision-Language Models

Pro-Bench: Prompt-Robust Open-Vocabulary Visual Grounding Across Real-World Heterogeneous Environments

Proactive Defense: Hardening Code Pipelines and CI/CD Infrastructure

Prompt, Probe, Train, or Annotate? Single-camera sports video understanding in amateur settings

Provably Complete Generalized Planning with LLMs

PSA: llama.cpp -cram should be increased for agentic workflows (default is 8192)

Quantifying the Occult: A Comparative Study of Hindu and Buddhist Deities Using Machine Learning Methods

Quantum score matching with applications to learning thermal states

Qwen-3.8-27B is good enough that I stopped using API

Qwen Image 2.1 masked inpainting: working crop-and-stitch graph + the wiring and prompt differences from Flux

Qwen3.8-Flash-Next on 12GB VRAM - 65 tokens per second

Qwen3.8-Flash-Next, on 5090+64gb, with Llama.cpp - Seems to not use ram?

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]

Reachable Global Optimization in AI Systems: How Global Is Global?

Realize What Matters: Principled Context Representation for Large-Scale Reasoning

Recognized but Not Produced: A Generation Benchmark for Culturally Specific Kinship Terms