Tag
Pathway's 150M-parameter BDH-CQ model achieves 29.5% on ARC-AGI-1 at a record-low cost of $0.0007 per task, using recurrent memory and latent reasoning instead of long token chains. The architecture may be the breakthrough Andrew Curran teased, with OpenAI researcher Lukasz Kaiser as an investor and adviser.
This paper presents ZeroLock, a backpropagation-free algorithm for concurrent memory-efficient LLM training that decouples model updates into independent chunk updates, reducing memory usage by 26.5% and improving throughput by 4.9% compared to BP-based baselines.
1jehuang launched Jcode, an open-source terminal coding agent written in Rust that claims 20x better memory efficiency than Claude Code, allowing dozens of agents to run in parallel.
Multiverse Computing announces a paper on making LLM knowledge distillation cheaper via offline top-K logits and a fused chunked KL loss, cutting VRAM usage for distillation at scale.
This paper introduces MESH, a memory-efficient Sinkhorn-based optimizer for Mixture-of-Experts (MoE) training that restores temporal momentum without storing full optimizer state, reducing memory by 62.5% while maintaining competitive evaluation loss compared to AdamW.
This paper investigates memory requirements for tabular foundation models like TabPFN and shows that model compression (e.g., INT4 quantization) can reduce memory footprint up to 7.6x with minimal accuracy loss, improving practical deployment efficiency.
WASTE is an open-source inference engine that runs models larger than available RAM by storing expert weights on NVMe. It demonstrates running Kimi K3, a 2.78T parameter MoE model, on a MacBook Pro with 64GB unified memory.
SkewAdam is a tiered optimizer that reduces memory usage for MoE state by 97%, enabling a 6.7B MoE model to fit on a single 40GB GPU.
SALT is an open-source tool that compresses long documents into a fixed-size plain-text prompt for LLMs, using a keyword trie to avoid theme collapse and efficiently select informative sentences under a token budget.
Qbix Server is a pure PHP web server that handles 10x more concurrent PHP requests on the same hardware compared to Nginx+PHP-fpm by forking workers after loading classes to share memory via copy-on-write, reducing per-worker memory from 30-60MB to ~5MB.
SkewAdam is a novel optimizer for mixture-of-experts models that tier allocates optimizer state across backbone, experts, and router, reducing memory footprint to 2.6% of AdamW while achieving better validation perplexity in controlled comparisons.
Unlimited-OCR, a new open-source OCR model from a Chinese company, solves the memory growth issue common in AI OCR tools by keeping memory usage flat regardless of document length, enabling single-pass reading of dozens of pages at 32K context. It's MIT-licensed, 3B parameters, multilingual, and already popular on GitHub.
CARE-LoRA proposes a compressed activation reconstruction framework to reduce memory consumption during LoRA fine-tuning by leveraging low-rank projections. It achieves competitive performance with reduced memory footprint.
StickyMoE proposes a differentiable routing consistency loss that encourages adjacent tokens to activate the same experts in MoE models, reducing expert-swapping overhead and cache misses during inference on edge devices by up to 3.92× while improving perplexity.
This tweet announces DeepSeek Unlimited OCR, an AI model that reads entire books in one pass with flat memory usage, achieving a 93% benchmark score and sub-0.11 error rate on 40+ pages.
This paper empirically shows that the gradient's top-r subspace in low-rank training methods like GaLore is non-identifiable beyond a small reproducible core, with estimator noise dominating apparent rotations. It analyzes the implications for optimizer state transport and introduces LDAdam, which outperforms GaLore in perplexity.
FourTune proposes a fully 4-bit quantization framework (W4A4G4) for efficient post-training of diffusion models, using a triple-branch hybrid pipeline and custom fused kernels to reduce memory by 2.25× and increase throughput by 2.27× on 12B FLUX.1-dev without quality loss.
China open-sourced Unlimited OCR, an OCR model based on DeepSeek OCR that can transcribe entire books in a single pass with constant memory usage, achieving 93% on standard parsing benchmarks.
Introduces MEET, a memory-efficient E(3) equivariant transformer for full-atom peptide design, integrated with a VAE and latent diffusion pipeline to achieve linear memory scaling and improved generation quality.
GRASP proposes a method for multi-source transfer learning that sequentially merges source models into a single target model with constant O(1) memory usage, using gradient-based parameter alignment to avoid negative transfer. Experiments show it outperforms ensemble methods while being much more memory-efficient.