AI developer tools · What shipped

Qwen3.8 Max leads the Vision2Web benchmark for converting visual references into functional web implementations with a score of 69.0%, followed by Qwen3.8-27B at 62.9%. The benchmark was updated August 17, 2026, with 2 multimodal models currently evaluated on BenchLM's display-only reference benchmark that measures screenshot-to-web task performance.

LTX-2 is an open-source DiT-based audio-video foundation model available on GitHub with multiple inference pipelines for video generation. The release includes LTX-2.5 with a 22B distilled transformer variant, separate modular weights for VAEs and text encoders, and optimization options including FP8 quantization and attention backends (FlashAttention 3/4, PyTorch SDPA). Multiple production-ready pipelines are provided: DistilledPipeline for fast inference (8 steps), DFRPipeline for detail-fidelity rendering with spatial upsampling, two-stage variants for higher quality, and specialized pipelines for video-to-video, audio-to-video, and lip-sync applications. The framework supports HDR/EXR output, gradient estimation for reduced inference steps, and includes training tools for LoRA and IC-LoRA fine-tuning through the ltx-trainer package.

DeepSeek launched DeepSeek-V4-Pro-0813, reporting benchmark improvements including 62.7 on DeepSWE (up from 12.8) and 87.9 on Terminal Bench 2.1, positioning it for autonomous AI agent and software engineering tasks. The company released DeepSeek Harness (dsh) v0.1, an open-source MIT-licensed developer tool built on modular architecture as an alternative to Claude Code for orchestrating autonomous coding agents. DeepSeek also implemented API billing changes moving from flat rates to peak/off-peak pricing with costs increasing 50-1,100% depending on model and usage time, though V4 Pro output remains priced at $3.96 per million tokens during peak hours—still below many Western competitors.

DeepSeek released DeepSeek Harness v0.1 in developer preview under the MIT license as an open source agent runtime framework. The tool uses the Cordis plugin architecture making models, tools, skills, sessions, sandboxes, storage, loops, scheduling, and UI all swappable components configurable without modifying source code. It offers four runtime modes (Standard, Code, Minimal, Creator) and maintains an append-only session log capturing every context injection for full run traceability, replay, and fork capabilities. The framework supports multiple model providers including DeepSeek, Anthropic, OpenAI, Bedrock, Vertex, Azure, and OpenAI-compatible endpoints, with deployable via npx as dsh or from source using pnpm, and includes a Python SDK for Python 3.10+.