Sewon Min's Preprint Shows Smaller Batches Beat Overfitting on Repeated Training Data
New research finds smaller batches and tuned recipes dramatically improve pretraining when you have more compute than unique data.
20 stories
New research finds smaller batches and tuned recipes dramatically improve pretraining when you have more compute than unique data.
A 0.3B EmbeddingGemma 2 fine-tune turns workflow states into typed, calibrated answers to yes/no, multiple-choice and ranked-level questions.
Sakana AI's MASS method lets a single 27B model act as its own team, judge, and teacher, boosting score-per-token by up to 1.6x on research benchmarks.

A free local kit turns a single character illustration into a rigged VRM avatar by splitting the work between you, Tripo, and a Codex agent.
A refusal-removed build of Qwen3.8-Flash-Next transplants 144 tensors into an already-quantized GGUF without recomputing a single learned value.
Tov is a new TypeScript-flavored language that compiles to native binaries, outpacing Rust on throughput, memory, and build times while giving agents machine-readable fixes.
A 7.6M-parameter Turkish text-to-speech model runs 50x faster than real time on a laptop CPU, with the fewest character errors among 15 Turkish systems benchmarked.
A community-maintained repackaging of Alibaba's Wan 2.1 video model lets 6GB consumer GPUs generate 480p clips through the WanGP toolbox.
Switzerland's fully open Apertus family grows eyes and ears, with a 262K context window, optional thinking mode, and Apache 2.0 weights for everything.
AutoTrust AI Lab released GEV-26B-Decide, a Gemma-4 based decision model that outputs calibrated probabilities and thinks only when uncertain.
A new sparse-view 3D asset pipeline fuses reconstruction and generation in a shared coordinate space, with code and weights now public.
Zed's multiplayer agent workspace Delta rolls out @mentions, pushing OS notifications and routing tagged conversations into a per-user Inbox.
The compact RepVGG-A0 checkpoint is live on the Hugging Face Hub via timm, offering a 9M-parameter plain-conv backbone trained on ImageNet-1k.
Google DeepMind's TAPNet repository bundles every generation of their point tracker, from TAP-Net to TAPNext++, in one open-source hub.
A free, open-source Mac app uses Google DeepMind's new multimodal embedding model to search inside photos, PDFs, audio, video and code locally.
Intfloat posted a deprecation notice on the original E5-base embedding model, telling users to migrate to E5-base-v2 for better performance.
Anthropic's free, nine-chapter prompt engineering course on GitHub has crossed 38,000 stars, offering hands-on Jupyter notebooks with graded exercises against the Claude API.
A C++ port with pre-quantized GGUF weights brings the ACE-Step 1.5 music foundation model to CPU, CUDA, Metal and Vulkan backends.
An open-source LangGraph agent swarm runs full red-team kill chains inside a Kali sandbox, scoring 98% on XBOW's validation benchmarks.
Anthropic opened OSS Scanner, a free opt-in service that runs its strongest models against critical open-source repos and emails raw, unreviewed vulnerability reports to maintainers.