KaniTTS2 — Open 400M TTS with Voice Cloning on 3GB VRAM
KaniTTS2 is a 400M-parameter open-source TTS model that runs voice cloning on just 3GB VRAM with full pretraining code released.
archive
Columns · Page 20
356 · Page 20
Training an LLM on CPU in 1.2 Hours
How FlashLM v3 trained a 13.6M-parameter LLM on CPU alone in 1.2 hours with MatMul-Free ternary-weight architecture, and its implications for edge AI.
Does AGENTS.md Actually Work? The First Empirical Study
The first empirical study evaluating AGENTS.md effectiveness has been published. We analyze its impact on coding agent success rates and inference costs.
AI Self-Generated Skills Are Useless
SkillsBench proves AI agents cannot author useful skills for themselves. Across 7,308 trajectories, self-generated skills showed zero benefit while human-curated skills improved performance by 16.2pp.
FunctionGemma 270M — 90-97% Tool Calling in a Tiny Model
Analysis of how fine-tuning FunctionGemma 270M improved multi-turn tool calling accuracy from 10-39% to 90-97%, matching a 120B teacher model. More evidence that scaling isn't everything.
4 of OpenRouter's Top 5 Models Are Open Source
Four of the top five most-used models on OpenRouter are open source (Qwen3-Coder, DeepSeek R2, MiniMax M2.5, etc.).
Qwen 3.5 Goes Bankrupt on Vending-Bench 2
Qwen 3.5, a top performer on standard benchmarks, goes bankrupt on Vending-Bench 2's vending machine simulation. Exploring the blind spots of benchmark-driven AI evaluation.
Claude Code with Local Models Triggers Full Prompt Reprocessing
Analyzing the full prompt reprocessing issue when running Claude Code with local LLMs. Learn about KV cache invalidation mechanics and developer tool design lessons.
Heretic 1.2: 70% VRAM Reduction via Quantization and MPOA Explained
Heretic 1.2 is here with 4-bit quantization cutting VRAM usage by up to 70% and MPOA delivering higher-quality abliteration. A deep dive into the latest cost-saving techniques for local LLM operations.
Karpathy: AI Training Costs Drop 40% Per Year
Karpathy's analysis reveals AI model training costs fall 40% annually. We examine the structural factors — hardware evolution, algorithm efficiency, and data pipeline optimization — and their industry impact.
How to Run Qwen3-Coder-Next 80B on 8GB VRAM
Analyzing quantization and lazy loading techniques to run an 80B parameter coding AI model on consumer 8GB VRAM GPUs. Exploring the practicality and limitations of local LLM coding.
Building SQLite with an AI Swarm
Six AI agents (Claude, Codex, Gemini) built a 19,000-line Rust SQLite clone in parallel. Analyzing the real costs of multi-agent coordination and task division.