AI Open Source Radar — July 26, 2026

Data source: Trendshift daily (freshness verified — top live-mention snowflake ID decodes to 2026-07-25 01:27 UTC, updated versus the prior run). The primary metric is recent (yesterday/today) star velocity, using Trendshift’s daily rank as a proxy. Yesterday’s featured repos have mostly decayed (block/buzz 1→5, ego-lite 5→10, and img2threejs / open-code-review / palmier-pro off the board), so this is not a stuck list. Four fresh risers today.

amap-cvlab/ABot-World — Trendshift Daily #1 (New 2026, surging)

A real-time interactive “world model” from Alibaba’s Amap CV Lab. On a single desktop GPU (NVIDIA RTX 5090) it delivers an endless, action-conditioned world rollout at 720p, 16 FPS, 1.2s latency, and 19GB of memory. The core technique is “LongForcing” training, which pushes past fixed video-length limits by expanding the world with new scenes and dynamics mid-rollout — no scene lock-in and no prompt switching — running stable continuous inference for over an hour (far beyond the typical one-minute limit of similar products). Released on July 9, it shipped the causal student model ABot-World-0-5B-LF, inference code, a local Gradio demo, and the online “ABot World Studio” playground, with a 500-hour action-annotated training dataset slated to be open-sourced.

Single desktop GPU (RTX 5090): 720p, 16 FPS, 1.2s latency, 19GB memory
Infinite, action-conditioned world rollout beyond fixed video length
LongForcing training expands scenes and dynamics mid-rollout, avoiding scene lock-in
Stable continuous inference for over an hour
Released Jul 9: ABot-World-0-5B-LF model, inference code, local Gradio demo, ABot World Studio; 500-hour annotated dataset to be open-sourced

Practical / customer angle: attach a real-time interactive world to games, simulation, robotics synthetic data, or autonomous-driving scenario generation — all on a single GPU.

Tags: #WorldModel #InteractiveVideo #RealTimeSimulation #ActionConditioned #SingleGPU #RTX5090

URL: https://github.com/amap-cvlab/ABot-World

andrewyng/openworker — Trendshift Daily #2 (New 2026, surging)

An MIT-licensed, local-first “desktop AI coworker” released by Andrew Ng on July 23 that aims to return finished deliverables (polished documents, calendar updates) rather than chat replies. The stack is a Tauri 2 + React shell over a local Python FastAPI agent server built on aisuite, so the agent loop, conversations, connector tokens, and model keys all stay on your machine. Model access is bring-your-own-key across 30 curated tool-calling models plus fully local Ollama, and it integrates with files, Slack, calendar, and other everyday tools to actually get work done. On safety, a typed risk engine (read / write_local / exec / external) gates every action across five permission modes, forcing user approval before high-risk operations run.

Local-first: agent loop, conversations, connector tokens, and model keys stay on your machine
Stack: Tauri 2 + React shell over a local Python FastAPI agent server (aisuite)
Bring-your-own-key across 30 curated tool-calling models, plus fully local Ollama
Integrates with files, Slack, and calendar to return finished deliverables
Typed risk engine (read / write_local / exec / external) with five permission modes; confirms before high-risk actions

Practical / customer angle: automate repetitive work — drafting documents, organizing schedules — into completed outputs, locally, without sending data outside.

Tags: #LocalFirst #AICoworker #BYOK #Ollama #TauriApp #PermissionGating

URL: https://github.com/andrewyng/openworker

ifixai-ai/iFixAi — Trendshift Daily #3 (New 2026, surging)

An open-source diagnostic for AI misalignment that probes the gap between what an agent was designed to do and what it actually does — the kind of quiet failure that can look perfect on every dashboard. It runs 45 inspections in total (32 graded core plus 13 extended) and returns a letter grade in under five minutes, grouped into five misalignment categories: fabrication, manipulation, deception, unpredictability, and opacity. The extended checks reach into frontier risks such as sabotage, sandbagging, and oversight evasion. It is provider-agnostic (OpenAI, Anthropic, Bedrock, Azure, Gemini, and more), can be dropped into CI to run repeatedly, and produces a content-addressed manifest for bit-identical replay. It is backed by iMe, a deterministic alignment runtime enforcing six constitutional rules through a six-stage pipeline with no LLM in the decision path.

45 inspections: 32 graded core + 13 extended
Five misalignment categories: fabrication, manipulation, deception, unpredictability, opacity
Extended checks cover frontier risks: sabotage, sandbagging, oversight evasion
Provider-agnostic (OpenAI, Anthropic, Bedrock, Azure, Gemini…), CI-friendly
Letter grade in under 5 minutes; content-addressed manifest for bit-identical replay
Backed by iMe: deterministic runtime, six-stage pipeline, six constitutional rules, no LLM in the decision path

Practical / customer angle: run it as a pre- and post-deployment regression check to catch misalignment before regulators or customers do.

Tags: #AIAlignment #AgentEvaluation #RedTeaming #DeterministicRuntime #CIIntegration #Reproducibility

URL: https://github.com/ifixai-ai/iFixAi

pax-beehive/paxm — Trendshift Daily #4 (New 2026, ~390⭐)

A provider-neutral memory layer that gives coding agents persistent memory, written in Go (Apache-2.0). It works across major coding agents and MCP-compatible clients — Codex, Claude Code, OpenCode, and Pi — so context, decisions, and preferences carry over even as sessions change. Its differentiators are neutrality (not tied to a specific vendor or IDE) and a lightweight, single-Go-binary orientation, which stands out in the recently crowded agent-memory category (cavemem, agentmemory, codebase-memory-mcp, and others) through a “provider-agnostic + MCP-first” positioning. It first entered Trendshift’s Go daily list at #4 on July 22, then climbed to #4 across all languages today.

Persistent, provider-neutral memory for coding agents
Works across Codex, Claude Code, OpenCode, Pi, and MCP-compatible clients
Written in Go (Apache-2.0), lightweight single-binary orientation
Carries context, decisions, and preferences across sessions
Provider-agnostic + MCP-first positioning in a crowded agent-memory category

Practical / customer angle: teams juggling multiple coding agents can share one memory layer and cut down on re-entering context.

Tags: #AgentMemory #MCP #GoLang #CrossAgent #CodingAgents #ContextPersistence

URL: https://github.com/pax-beehive/paxm