AI Wiki — Knowledge Base

R9V: A designer set of kernels I've been working on for R9700s/RDNA4. Qwen3.8-Flash-Next Unsloth IQ4_XS (w/ TP on 2 R9700s, MTP, SSD n-gram, 128k ctx, vision): TG256 of 78 tok/s (~3x increase), PP8192 of 1510 tok/s (~30x increase).

TL;DR: Ninfer/DS4 but for RDNA4 Highly custom kernels built for RDNA4, applied to vLLM-Radiance to greatly improve Qwen3.8 Flash Next speeds. This is mostly for dual R9700s with preferably 48GB of RAM

model-releasesr-localllama

31 Aug 2026

Search: “models”

Pretraining and Distillation Matter More Than Architecture Family for Label-Free Single-Cell Classification

SEA-SpeechBench: A Large-Scale Multitask Benchmark for Speech Understanding Across Southeast Asia

Semigroup-JEPA: Latent Dynamics Consistency for Zero-Shot Physics Generalization

Adapting Neural Robot Dynamics on the Fly for Predictive Control

Ambient @ EgoProactive 2026 : Proactive Egocentric Assistance with Visually Grounded Supervision

DC-Gen: Post-Training Diffusion Acceleration with Deeply Compressed Latent Space

Hyperparameter Scaling Laws Across MoE Sparsity

MOLE: Detecting Insider Threats in AI Agents

STSG-VQA: Evidence-Grounded Temporal Question Answering from Surgical Spatio-Temporal Scene Graphs

Surveillance plagiarism by OpenAI

We benchmarked DeepSeek V4.1 Flash by @deepseek_ai . It reached 98% of GPT-6 Astra’s score at 1.4% of the cost on everyday design tasks base…

How many agents can 2×4090 actually run at once? Three weeks of llama.cpp concurrency data — soft cap 5 @ 64k, hard cap 9, and why.

I built Infercat: Share your local AI with friends over an encrypted p2p tunnel. Connect with an invite code from anywhere.

Qwen3.5 0.8B on CPU

LLM regression in reading comprehension?

The gap has closed, open source will win

Even Qwen3.8 followed the instruction inside my translation data, and Gemma 4 beat the translation specialists I tested

Spectral characteristics of autoencoder parameters as a vector representation of data

DocHop: Benchmarking Out-of-domain Multi-hop Reasoning in Information-Dense Documents

How Fast Do Agents Rot? An Empirical Study of Long-Horizon Degradation in LLM Agents for Production Decision-Making

Untangling the Mechanisms of Misleading Context in Medical Question Answering

Claude's new system prompt really doesn't want to reproduce song lyrics

Compositional Machine Design as Program Synthesis with LLMs

Do General NLP Embeddings Capture Ontological Reasoning?

EvoFlint: An Evolutionary Atlas of Multi-Turn LLM Vulnerabilities

Invalidation Contracts for Cross-Episode Agent Memory

LLM-Driven Autonomous Vehicles Inherit Human Driver Biases in Pedestrian Yielding: Results and Implications From A New Benchmark

Post-Training Science for Supervised Fine-Tuning

SOVER: Formal Certification of Optimization Reformulations via LLM-Assisted SMT Verification

AcrossWAM1.0:A Modular Latent World-Action Stack for Compact Robot Policies

ACTD: Anchor-Based Cross-Tokenizer Distillation with Residual Regularization

Generating Workflow DAGs from Natural Language with Non-Reasoning LLMs

GeoAgent: Evaluating VLM Geolocalization Through Embodied Navigation

Learning to Follow In-Context Watermark Instructions via Self-Distillation

Measurement Validity in LLM Cultural Alignment

MedSegBenchmarker: A Raw-Count-First Framework for Controlled 2D Medical Image Segmentation Benchmarks

Real-Time Deadlines Reveal Fragile Temporal Adaptation in LLM Strategic Dialogues

SciOrch: Learning to Orchestrate Expert LLMs for Solving Frontier Multimodal Scientific Reasoning Tasks

Self-Specialized Teachers for Domain Post-Training

SUP-MIMIC: A Multi-Task Clinical Diagnosis Benchmark for Evaluating LLMs' Robustness to Contradictory Evidence

The Safety Relay in Roleplay Jailbreaks: A Component-Resolved Causal Analysis of Harm Recognition and Refusal

Where Induction Runs Out: Description-Length Difficulty and the Memorisation Gap in Integer-Sequence Benchmarks

Characterization of Request and Token Energy Costs for LLM Inference Workloads on GPU Platforms

Memorization Is Not Extraction: Tight Differential-Privacy Bounds and Audit Blind Spots

R9V: A designer set of kernels I've been working on for R9700s/RDNA4. Qwen3.8-Flash-Next Unsloth IQ4_XS (w/ TP on 2 R9700s, MTP, SSD n-gram, 128k ctx, vision): TG256 of *78 tok/s* (~3x increase), PP8192 of *1510 tok/s* (~30x increase).

Speculative Probing: LLM Monitoring at Speculative-Decoding Cost

Qwen 3.8 27B at 50 tok/s with 100k Context on a 16GB GPU! (beellama.cpp)

4DSynth: Controllable Procedural World Synthesis for Dynamic Embodied Simulation

R9V: A designer set of kernels I've been working on for R9700s/RDNA4. Qwen3.8-Flash-Next Unsloth IQ4_XS (w/ TP on 2 R9700s, MTP, SSD n-gram, 128k ctx, vision): TG256 of 78 tok/s (~3x increase), PP8192 of 1510 tok/s (~30x increase).