DeepSeek's DSpark speculative decoding technique, which combines parallel drafting with a lightweight sequential component, can improve local LLM generation speed on the same GPU, with DeepSeek report…
Zhipu AI's open-weight GLM 5.3-Flash, released under an MIT license, processed roughly 23 trillion tokens in its first week on OpenRouter, becoming the platform's most-used model. With 320B parameters…
A proof-of-concept repository, 'photonic-mesh-fng-router', proposes a hardware-native optical timing control plane that uses GPU register-level fluid dynamics to eliminate optical transmission jitter …
Cursor Research released Mixture-of-Kittens (MoK), a CUDA kernel that fuses mixture-of-experts computation and communication, delivering a 41% training throughput increase on Nvidia GB300 NVL72 system…
A developer reports cutting Kimi K3 costs in OpenCode by switching to the k3-256k model, which consumes half the quota of the standard K3, and by choosing the right provider such as Novita.ai. The All…
An experiment by an unnamed researcher found that retrying failed requests in LLM benchmarks introduces selection bias, shortening average response lengths by 19.2% and reducing long-form outputs by u…
DeepSeek, the Hangzhou-based AI startup, is building its first known data center in Ulanqab, Inner Mongolia, according to job postings dated April 2, 2026, as it prepares to launch its trillion-parame…
Every major LLM vendor has replaced numeric thinking-token budgets with discrete effort levels, converging on a small enum instead of a number. Anthropic now rejects budget_tokens with a 400 error on …
Harness Engineering built a fully automated coding loop inside OpenCode using DeepSeek-V4-Pro and DeepSeek-V4-Flash, enabling low-cost agentic code iteration. The loop, inspired by Andrew Ng's breakdo…
A developer spent 107 days running 9 autonomous AI agents in a physical fitness studio and discovered that structural constraints matter more than model intelligence. The system, built on DeepSeek-V4 …
Chinese AI firm DeepSeek open-sourced DSpark, a speculative decoding system that accelerates large language model inference by up to 85% without altering output quality, releasing it under the MIT lic…
Fastllm, a C++ LLM inference library, now supports running DeepSeek-V4 and the full DeepSeek R1 671B model on a single GPU with just 10GB VRAM. The library is compatible with Nvidia, AMD, and domestic…
DeepSeek released DSpark on June 27, a speculative decoding framework that accelerates per-user generation speeds by 60% to 85% on its DeepSeek-V4 Flash model and 57% to 78% on the Pro variant. The fr…
DeepSeek open-sourced DSpark, a speculative decoding framework that accelerates per-user generation for DeepSeek-V4 by 57–85% over the MTP-1 baseline. The framework pairs a parallel draft backbone wit…
AMD introduces DP Attention and Two-Batch Overlap (TBO) optimizations for DeepSeek-V4 inference on MI355X GPUs, using a coordinated prefill scheduler called PrefillDelayer to reduce padding waste and …
A developer created a plugin called 'observer' for OpenCode that enables the DeepSeek-V4 language model to read images by calling a multimodal agent, allowing it to interpret error screenshots, charts…
DeepSeek-V4's million-token context capability stems from a hybrid attention architecture that compresses context before KV storage, reducing cache pressure. Together's early bring-up on NVIDIA HGX B2…
DeepSeek-V4 introduces a new architecture using hybrid attention mechanisms—Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA)—to drastically reduce the computational cost and me…
DeepSeek-V4 shifts the focus from simply offering a million-token context window to making that window practically usable through efficient attention mechanisms. Its architecture employs hybrid compre…