Notes on models

Showing notes about models

Topics

9 notes on this topic

ai (21)agents (14)models (9)memory (7)architecture (6)research (4)scaling (4)systems (4)

Show 66 more topics

trajectories (4)efficient-models (3)forecasting (3)hugging-face (3)infrastructure (3)open-source (3)physics (3)alignment (2)compute (2)image-generation (2)local-ai (2)open-weights (2)oxygen-ai (2)papers (2)reasoning (2)self-hosting (2)verification (2)agency (1)ai-infrastructure (1)attention (1)coding (1)commons (1)cryptography (1)css (1)decentralization (1)desktop (1)developer-tools (1)diffusion (1)efficiency (1)frontend (1)gemma (1)geometry (1)hardware (1)hermes (1)image-editing (1)inference (1)interfaces (1)introduction (1)kwaipilot (1)learning (1)llm (1)long-context (1)microsoft (1)minimax (1)multimodal (1)neuroscience (1)nextjs (1)nous-research (1)on-device (1)ornith (1)plugins (1)policy (1)privacy (1)react (1)recursive-self-improvement (1)reinforcement-learning (1)safety (1)security (1)singularity (1)swarms (1)swe-bench (1)tools (1)trust (1)tutorial (1)ui (1)vision (1)

airesearchpapers

Looped Transformers: Depth Without New Parameters

A working note on looped transformers: reuse one layer stack across passes to buy effective depth. Nanbeige4.2-3B is the latest compact example, with author-reported agent scores that beat much larger models.

01 · Sep 7, 2026Read

aiagentscoding

KAT-Coder-V2.5: Infrastructure Beats Scale for Agentic Coding

KwaiKAT’s KAT-Coder-V2.5 comes within a few points of Opus 4.8 on repository SWE, tops PinchBench, and the open KAT-Coder-V2.5-Dev (35B MoE / 3B active) leads same-scale peers. The paper’s real story is environments, trajectories, and harness-aware RL—not parameter count.

02 · Jul 26, 2026Read

aimodelsimage-generation

Mage-Flow: Open Image Generation and Editing from Microsoft Research

Microsoft Research open-sourced Mage-Flow, a compact 4B stack for text-to-image and instruction editing. I ran the full family locally, and a hard, structured prompt produced this wizard key art—proof that open image models are getting seriously good.

03 · Jul 25, 2026Read

aiopen-sourceopen-weights

The Open-Weights Letter and the Labs That Stayed Off the Page

Read the July 24 open-weights statement. Then look at who would not put their name on it—Anthropic, Google, OpenAI, and xAI—each for different reasons. The letter is about diffusion. Their silences are about control.

04 · Jul 25, 2026Read

aimodelsdiffusion

DiffusionGemma: The Developer Guide

Google's experimental DiffusionGemma (26B MoE on the Gemma 4 backbone) replaces autoregressive token-by-token generation with diffusion-based parallel decoding on 256-token blocks, delivering up to 4x faster inference, bidirectional context, and built-in self-correction while remaining practical to serve and fine-tune.

05 · Jun 11, 2026Read

aimodelsimage-generation

Bonsai Image 4B: Local Ternary Image Generation on an M4 Pro

A long-form tutorial for running PrismML's Bonsai Image 4B locally with the ternary MLX backend, higher-quality 4-step previews, source links, and honest M4 Pro measurements.

06 · Jun 7, 2026Read

aiagentsmodels

Holo3.1 and Desktop Vision Agents

H Company’s Holo3.1-35B-A3B is less interesting as another model leaderboard entry than as a signal about local computer-use agents, UI grounding, and efficient active-parameter architectures.

07 · Jun 6, 2026Read

aimodelsmultimodal

Gemma 4 12B Unified

Google's open-weight Gemma 4 12B-it folds text, image, and audio into a single model with no separate encoders, runs on a 16GB Mac after 4-bit quantization, and posts benchmarks well above its parameter count.

08 · Jun 4, 2026Read

aimodelsresearch

HRM-Text-1B and Efficient Pretraining

Sapient's HRM-Text-1B is a useful signal: architecture and training objective can still move the efficiency frontier, not just model size.

09 · May 21, 2026Read