AI news — August 1, 2026
The 20 top stories PulseAugur surfaced that day, ranked by signal across labs, papers, and developer communities.
-
Sber releases open-weights GigaChat 3.5 Ultra 432B model
Sber has released GigaChat 3.5 Ultra, a 432B parameter model with open weights under the MIT license, focusing on coding, agents, and long context. While the open weights remove licensing barriers, users must still verify runtime compatibility, hardware requirements, and actual …
-
OpenAI unveils Astra model family for multi-agent collaboration
OpenAI is developing a new model family named Astra, designed to enable multiple AI agents to collaborate on complex tasks over extended periods. CEO Sam Altman has showcased Astra to policymakers, and the company is considering releasing it as either GPT-6 or a new variant of G…
-
GigaChat 3 Ultra API access opens for individuals, prompting developer testing
GigaChat 3 Ultra has been made available to individuals via API on a freemium basis, allowing for testing without a corporate account. This release includes updates to API endpoints and model information, aiming to simplify integration. Developers are advised to conduct thorough…
-
Supabase releases open-source benchmark for AI coding agents · 2 sources tracked
Supabase has released an open-source benchmark and framework called Evals to evaluate AI coding agents. The tool tests agents like Claude Code, Codex, and OpenCode on real-world Supabase tasks, such as schema creation and debugging, using containerized environments for realistic…
-
Kimi K3 challenges 'open equals cheap' AI model pricing
Kimi K3, an open-weight model, is challenging the long-held assumption that open-weight models are inherently cheaper than proprietary APIs. This model matches the pricing of Anthropic's Claude Sonnet 5 while claiming performance comparable to Claude Opus 4.8. This development n…
-
LiteLLM and LangGraph unify 176 LLM APIs for seamless switching
A new approach using LiteLLM and LangGraph has been developed to unify the interfaces of over 176 large language models, including those from OpenAI, Claude, Qwen, and DeepSeek. This system addresses the significant challenge of high switching costs and code modifications requir…
-
DeepSeek V4 officially launches, promising enhanced capabilities and value
DeepSeek V4 has officially launched, introducing new capabilities and aiming for a competitive price point. This release positions the model as a cost-effective option in the AI market. The announcement comes amidst broader news of record-breaking heat in the US and adjustments …
-
AMD releases open Instella-MoE-16B LLM with 2.8B active parameters
AMD has released Instella-MoE-16B-A3B, an open-source Mixture-of-Experts language model. This model features 16 billion total parameters but only activates 2.8 billion per token, utilizing architectural innovations like Gated Multi-head Latent Attention and FarSkip-Collective co…
-
China's Moonshot AI releases Kimi K3, largest open-weight model
Moonshot AI, a Beijing-based lab, has launched Kimi K3, an open-weight AI model boasting 2.8 trillion parameters, making it the largest of its kind. The model demonstrated advanced capabilities by designing a chip to run a smaller version of itself. Demand for Kimi K3 surged sho…
-
DeepSeek V4 Flash challenges top AI models with drastic price cuts · 1 source tracked
DeepSeek has launched a new coding model, V4 Flash, which rivals the performance of Anthropic's Claude Opus 4.8 on complex coding tasks. This release is part of a broader trend of AI models becoming increasingly commoditized, with DeepSeek offering its model at a fraction of the…
-
OpenAI's GPT-Red model finds AI vulnerabilities but doesn't guarantee agent safety
OpenAI has developed GPT-Red, an AI model designed to iteratively attack other AI models and generate adversarial data for their training. While GPT-Red demonstrates significant success in finding vulnerabilities, particularly in indirect prompt injection scenarios where it outp…
-
NVIDIA Transformer Engine tutorial details GPU acceleration for LLMs
NVIDIA's Transformer Engine is detailed in a tutorial that explains how to accelerate transformer workloads. The engine combines fused GPU kernels, BF16 computation, and hardware-aware FP8 execution. The tutorial covers installation, GPU capability detection for TE kernels and F…
-
DeepSeek V4 officially released, South Korea's exports surge on AI chip demand
DeepSeek has officially released its V4 model, showcasing new capabilities and aiming for a competitive price point. This release comes as South Korea's exports surged in July, largely driven by strong demand for memory chips and significant growth in semiconductor exports, whic…
-
DeepSeek V4 launches with new features, amid strong cargo growth at Zhejiang airports
DeepSeek V4 has officially launched, introducing new capabilities and aiming for a competitive price-performance ratio. This release follows a period where Zhejiang province airports handled nearly 40,000 tons of fresh goods in the first half of the year, a significant increase …
-
DeepSeek V4 launches as cost-effective AI, while Harmony Intelligent Driving sees strong delivery growth
DeepSeek V4 has officially launched, showcasing new capabilities and positioning itself as a high-value option in the AI market. The announcement comes alongside news of Harmony Intelligent Driving's monthly deliveries reaching 45,046 units, contributing to a year-to-date total …
-
DeepSeek V4 official version released with new capabilities
A new version of the DeepSeek V4 model has been released, showcasing enhanced capabilities and competitive pricing. This release positions DeepSeek V4 as a strong contender in the market, potentially sparking a price war. The announcement was made by 36Kr, a financial informatio…
-
DeepSeek V4 officially launches with new capabilities
DeepSeek V4 has officially launched, revealing new capabilities and positioning itself as a cost-effective option in the market. This release comes amidst news of potential record-breaking heat in the US for July and Ford's preparations for Chinese automakers entering the US mar…
-
Build a 10M-Parameter GPT-Style Language Model From Scratch
This article provides a comprehensive guide to building a small, GPT-style language model from scratch using PyTorch on a home computer. It details the process in five stages: creating a tokenizer, preparing a dataset for next-token prediction, implementing the Transformer archi…
-
Gemma 4 26B model runs in specialized 2GB resident memory config
A project called TurboFieldfare has demonstrated a specialized configuration of Google's Gemma 4 26B model that utilizes approximately 2GB of resident memory on Apple Silicon. This is achieved by streaming model experts from an SSD rather than keeping the entire model in RAM, a …
-
LangGraph framework enhances LangChain for stateful AI agent workflows
LangGraph, a framework built on LangChain, enables the creation of stateful, multi-step AI workflows modeled as directed graphs. It addresses limitations of linear pipelines by allowing agents to retry failed tool calls, branch based on LLM decisions, incorporate human review, a…