AI News · August 8, 2026

DeepSeek V4 Flash price jumps 56% as SGLang adds Kimi K3 support

62,787 stars today75 new projects+1,740 skills
By olud.ai editorial · built from our own tracking data, published daily

DeepSeek V4 Flash 0423 input price rose 56% to $0.14 per 1M tokens, the largest price move today. SGLang v0.5.17 shipped with day-0 support for Kimi K3, a 2.8T-parameter multimodal model. The open-source AI ecosystem gained 62,787 GitHub stars today across 10,000+ tracked projects, with 75 new projects entering the index. This week 2,375 new projects were added, plus 100 new Hugging Face spaces.

Releases of the day

SGLang v0.5.17 arrived with 582 pull requests from 194 contributors, adding native support for Kimi K3, a 2.8T-parameter multimodal LatentMoE model using 896 experts. ComfyUI v0.31.0 bumped its frontend package and fixed memory pinning on Linux. Pydantic AI v2.27.0 added xai_agent_count to XaiModelSettings and improved round-trip compaction. Phoenix arize-phoenix v19.19.1 fixed OTLP path issues in Mastra snippets and cleaned up endpoint variables. Llama.cpp b10327 fixed thread and block counts in quantized CUDA kernel launches. ROS 2 released the Lyrical Luth Patch Release 2 binary packages.

CComfyUI v0.31.0 What's Changed * Bump comfyui-frontend-package to 1.47.12 by @comfy-pr-bot in https://github.com/Comfy-Org/ComfyUI/pull/15244 * Don't pin too much memory on Linux systems with... SGLang v0.5.17 Highlights · *582 PRs from 194 contributors.* · Kimi K3 day-0 support: A 2.8T-parameter multimodal LatentMoE (896 experts, top-16, routed in a 3584-dim latent space) with a... Pydantic AI v2.27.0 What's Changed 🚀 Features * Add xai_agent_count to XaiModelSettings by @pydanty[bot] in https://github.com/pydantic/pydantic-ai/pull/7155 * Round-trip CompactionPart through... Phoenix arize-phoenix-v19.19.1 19.19.1 (2026-08-08) · Bug Fixes · * app: make the Mastra snippet's OTLP path append idempotent (10688a4) * close remaining endpoint-variable foot guns in Python and onboarding... llama.cpp b10327 CUDA: fix thread/block count in quantized cpy kernel launches (#26731) · * CUDA: fix thread/block count in quantized cpy kernel launches · * tests: add uneven block count cpy... ROS 2 release-lyrical-20260807 These are the binary packages for ROS Lyrical Luth Patch Release 2 (2026/08/07) · For runtime dependencies, see the binary package installation instructions. Your system must...

Pricing moves

DeepSeek V4 Flash 0423 input price increased 56% to $0.14 per 1M tokens. Kimi K2.6 input rose 2% to $0.58, and Qwen3.5-122B-A10B input rose 12% to $0.29. Larger declines included Qwen3 VL 235B A22B Thinking down 59% to $0.40, Ling-3.0-flash down 75% to $0.02, GLM 5.2 down 15% to $0.47, and Inkling Small down 10% to $0.45. Most models saw price reductions, while DeepSeek V4 Flash moved sharply higher.

▲ 56%DeepSeek V4 Flash 0423DeepSeek · input $0.09 → $0.14 per 1M tokens
▼ 15%GLM 5.2Z.AI · input $0.55 → $0.47 per 1M tokens
▲ 2%Kimi K2.6Moonshot AI · input $0.57 → $0.58 per 1M tokens
▲ 12%Qwen3.5-122B-A10BAlibaba · input $0.26 → $0.29 per 1M tokens
▼ 59%Qwen3 VL 235B A22B ThinkingAlibaba · input $0.98 → $0.4 per 1M tokens
▼ 10%Inkling SmallThinkingmachines · input $0.5 → $0.45 per 1M tokens
▼ 75%Ling-3.0-flashInclusionai · input $0.08 → $0.02 per 1M tokens

New models

SGLang v0.5.17 added day-0 support for Kimi K3, a 2.8T-parameter multimodal LatentMoE with 896 experts and top-16 routing in a 3584-dim latent space. This is the first inference framework to support the model.

Emerging projects

Auto-Empirical-Research-Skills collected 3,309 stars, offering 23,000 agent skills for empirical research across eight social science disciplines. Hope-agent reached 1,388 stars, a cross-device desktop AI agent in Rust with memory, persistent goal pursuit, and dynamic multi-agent orchestration, designed to run on NAS or cloud.

Auto-Empirical-Research-SkillsStata · 3,309 stars · 🔬 A curated collection of 23,000+ agent skills for empirical research across 8 social science disciplines.... hope-agentRust · 1,388 stars · 🦭 会记忆、能持续推进目标、会动态编排多 Agent 的跨端桌面 AI 助手，也可服务化常驻 NAS / 云端 | A cross-device desktop AI agent with memory, a…

Research pick

AgentOPSD proposes recursive self-distillation for agentic reinforcement learning. Standard RL with verifiable rewards builds trajectory-level advantage estimates but fails to credit the few critical decisions in long-horizon multi-turn tasks. AgentOPSD addresses this by recursively distilling advantages at the decision level.

AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement LearningZi-Han Wang, Zhengxi Lu, Zhiyuan Yao et al. · 73 community upvotes · Reinforcement learning (RL) with verifiable rewards constructs trajectory-level advantage estimates, yet it often fails to credit the few pivotal decisions that determine outcomes in long-horizon, multi-turn agentic tasks. Recent work intro

The bigger picture

Price volatility continues, with DeepSeek V4 Flash jumping sharply while most other models dropped. SGLang’s rapid support for a massive new model and weekly ecosystem growth of over 2,300 projects point to accelerating infrastructure and tooling for agentic and multimodal AI.

Source: olud.ai tracking of 10,000+ open-source AI projects, 300+ models and live provider pricing. Figures are measured, not estimated. All releases · Live pricing · Latest in AI

← All AI news