Techyon

DeepSeek-v4.1 Flash: Pushing the Limits of KV Cache Compression

DeepSeek-V4.1 Flash is a 552B parameter multimodal mixture-of-experts model engineered to maximize KV Cache compression for long-horizon agent workflows. To address the storage and communication bottlenecks of long conte…

→ View original source

I built a native Vulkan training backend for 143 modern Transformer architectures — no CUDA or PyTorch required

The author created a native Rust + Vulkan training backend for the Hierarchos project that enables training and inference across 143 canonical Transformer architectures without requiring CUDA or PyTorch. The backend also…

→ View original source

China's open-weight AI models are now just 4 months behind frontier US offerings, Mozilla report claims — models still lag in some benchmarks but are drastically cheaper to use

According to a Mozilla report, China's open-weight AI models are now roughly four months behind the leading US frontier models, though they still lag on certain benchmarks. However, these models are drastically cheaper t…

→ View original source

I made this: Anyworld — a local-first single/multiplayer text RPG where the DM is your own LLM

Anyworld is a browser‑based, local‑first text RPG that lets you run your own LLM as the Dungeon Master, supporting both single‑ and multiplayer modes. Inspired by AI Dungeon 2, it uses OpenAI‑compatible backends (tested …

→ View original source

Apple reportedly building server packed with M-series Ultra chips for AI

Apple is reportedly developing a server packed with its M-series Ultra chips for AI workloads, targeting a debut around 2029. This would be Apple's first enterprise server offering in several decades, signaling a move in…

→ View original source

The Other Half of the Memory Wall: Serving 35B MoEs from SSD with Trained Routing Prediction

Researchers have introduced Edge0, a streaming inference engine designed to mitigate the memory wall encountered when serving large Mixture-of-Experts (MoE) models on consumer-grade hardware. While 35B-class models requi…

→ View original source

CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents

CERA-MoA (Co-Evolving Router with continually learning Agents for Mixture-of-Agents) introduces an iterative reinforcement learning framework designed to synchronize query routing with the evolving capabilities of LLM ag…

→ View original source

ScienceIDE: Turning World's Scientific Codebase into Agent Learnable Environments

To address the "scientific experience bottleneck" caused by fragmented toolchains and implicit domain conventions in scientific software, researchers introduced ScienceIDE, an infrastructure designed to transform complex…

→ View original source

A Zeroth-Order Paradigm for LLM Preference Alignment

Researchers have introduced Comparison-based Preference Optimization (ComPO), a novel zeroth-order paradigm designed to align large language models (LLMs) with human preferences. Unlike traditional direct alignment metho…

→ View original source

Qwen3.8 Max (0902) scores 45 on the Artificial Analysis Intelligence Index, up 5 points in a month and back on top of China's leaderboard, nosing out GLM-5.3 (44.9) and Kimi K3 (43.8)

Qwen3.8 Max (0902) achieved a score of 45 on the Artificial Analysis Intelligence Index, marking a five‑point increase over the past month. This performance reclaimed the top spot on China's leaderboard, surpassing GLM‑5…

→ View original source

You can offload most of Qwen3.8-Flash-Next's KV cache to RAM with little decode slowdown

Researchers show that most of the KV cache for Qwen3.8‑Flash‑Next can be moved to system RAM with minimal impact on decoding speed, allowing models that barely fit in VRAM to run at full context length without KV‑cache q…

→ View original source

I spent hours going through 100+ page PDFs, so I built a tool that highlights exactly where the answer came from. It's now completely open-source.

The author created an open-source tool that highlights the exact passages within long PDFs used to generate answers, providing traceable citations for LLM responses. It was developed after manually reviewing 100+ page PD…

→ View original source