Show HN: KBlip – turns AI/LLM news across 100 sources into daily digest threads

Releases / Launches

New models, tools, and framework launches

📣 News — Moonshot AI releases Kimi K3 open weights: 2.8T-parameter MoE model with 1M context, 91.2% on SimpleQA [19]

🛠️ Tool — WISP streaming engine runs 2T+ MoE models like Kimi K3 on consumer hardware via 3-tier VRAM→RAM→NVMe SSD [6]

🛠️ Tool — Krasis runtime streams Ornith-397B at Q4 on single RTX PRO 6000 Blackwell 96GB, achieving 2,354 tok/s prefill [3]

🛠️ Tool — OpenClaude improved coding agent released on GitHub, supporting cloud APIs, gateways, and local models [21]

🛠️ Tool — World-Model-Optimizer open-source tool distills and routes models to cut costs by half [20]

🛠️ Tool — Minimax M3 support with Multi-Stream Attention merged into llama.cpp for local long-context inference [22]

🛠️ Tool — POCKET-35B agentic model runs on CPU at 59 t/s, fits on phones with GGUF quantization [23]

🛠️ Tool — ai-sage releases GigaChat Audio 10B, an audio-native LLM with temporal grounding [24]

🛠️ Tool — Open WebUI v0.11.0 ships full interface redesign with chat view, admin panel, and lighter typography [8]

🛠️ Tool — ASL V6 open-source AST red-teaming engine for Python AI agents released on GitHub [18]

Today

29 items

Tool Release

Vanara launches free tier for CI-verified Claude Code agent catalog

Vanara is a catalog of Claude Code agents that are verified via runnable scripts in public CI on every push, ensuring they work as claimed. The free tier is now open, allowing users to install agent packs like npx vanara install security-pack.

1 engagement1 sourceredditMon, Jul 27, 2026, 02:44 PM

Agent Release

Cursor research department tests parallel multi-agent swarm skill

A user shares insights from Cursor's research department on agent swarms, applying similar ideologies to a new parallel multi-agent swarm skill. The skill aims to improve efficiency by coordinating multiple agents concurrently.

1 engagement1 sourceredditMon, Jul 27, 2026, 02:36 PM

Tool Release

Krasis runtime runs Ornith-397B at Q4 on single RTX PRO 6000 Blackwell 96GB

A developer built Krasis, an MoE-focused runtime that streams large models through limited VRAM. It runs Ornith-1.0-397B at Q4 on a single RTX PRO 6000 Blackwell 96GB, achieving 2,354 tok/s prefill and ~20–24 tok/s decode by keeping experts in CPU RAM and dynamically managing VRAM residency.

23 engagement1 sourceredditMon, Jul 27, 2026, 01:59 PM

Agent Release

AI coding agent refactors 750k LOC app in 3 days with zero bugs

A case study from AI Sovereign Labs shows an AI coding agent autonomously refactoring a 750,000-line application in three days. The agent ran 31 verification passes, corrected 201 errors, and shipped code with zero bugs, regressions, or technical debt. The project was not a from-scratch rewrite or a clone of training data.

3 engagement1 sourcehackernewsMon, Jul 27, 2026, 12:28 PM

Model Release

XYZ AI Lab releases XYZ-Aquila-mini, an open-weight Deep Search agent

XYZ AI Lab released XYZ-Aquila-mini, an open-weight Deep Search agent post-trained from Qwen3.6-35B-A3B via an AI4AI pipeline. The model is a thinking model designed for bounded-exploration search tasks, with human-defined constraints and AI-driven failure diagnosis.

4 engagement1 sourceredditMon, Jul 27, 2026, 12:00 PM

Tool Release

WISP streaming engine runs 2T+ MoE models on consumer hardware

WISP is a C/CUDA streaming engine that enables running large MoE models like Kimi K3 (2.8T) and GLM-5.2 (744B) on consumer hardware via 3-tier streaming (VRAM→RAM→NVMe SSD). It features self-organizing LRU cache, MLA attention support, and same-family speculative decoding for 2.2-2.8x throughput. Verified working on Mixtral-8x7B.

1 engagement1 sourceredditMon, Jul 27, 2026, 11:20 AM

Tool Release

Community ports SGLang to V100 GPUs with custom FlashAttention for Qwen models

A developer forked SGLang to add support for NVIDIA V100 GPUs, writing custom TeilLang FlashAttention and integrating open-source marlin-v100 and ungated flashinfer for sm70. The fork enables running Qwen3.5/3.6 models with Dflash and Laguna S2.1 support on 4xV100 32GB NVLINK, achieving ~4000-6000 prefill tokens/s and ~100 tokens/s generation.

12 engagement1 sourceredditMon, Jul 27, 2026, 10:20 AM

Tool Release

Open WebUI releases v0.11.0 with full interface redesign

Open WebUI v0.11.0 ships a ground-up visual redesign of the entire user interface, including chat view, admin panel, narrower conversation column, lighter typography, and rearranged settings. The update aims to improve consistency and usability across all aspects of the open-source web UI for LLMs.

18 engagement1 sourcegithubMon, Jul 27, 2026, 09:30 AM

RSS

Model Release

NVIDIA releases Nemotron 3 Embed, a high-performance open embedding family

NVIDIA released the Nemotron 3 Embed family of open embedding models, claiming top performance on retrieval benchmarks. The models are available via NVIDIA NIM and a semantic search demo, aiming to improve RAG and agent workflows.

0 engagement1 sourcerssMon, Jul 27, 2026, 09:09 AM

Tool Release

Skill Router: local-first tool for searching large agent skill libraries without blowing up your context window

Skill Router is a local-first tool that indexes agent skill metadata into SQLite, keeping full content out of context until selected. It supports CLI, interactive shell, MCP stdio server, and an optional HTTP API.

1 engagement1 sourceredditMon, Jul 27, 2026, 07:55 AM

Tool Release

Agent Context Lens v0.2.0 adds AGENTS.md chain inspection for Codex

A developer released Agent Context Lens v0.2.0, a free MIT-licensed CLI tool that visualizes how Codex resolves layered AGENTS.md instructions. It helps users understand which instructions are active, shadowed, or outside the chain, aiding debugging of large repositories.

1 engagement1 sourceredditMon, Jul 27, 2026, 07:50 AM

Tool Release

User seeks llama.cpp GUI for Windows with fast model/system prompt swapping

A user reports that llama.cpp runs well but lacks a GUI for quickly swapping models and system prompts. Lm-studio offers these features but runs significantly slower with the same settings, likely due to memory issues. Llama-swap handles model swapping but not system prompts.

8 engagement1 sourceredditMon, Jul 27, 2026, 07:41 AM

Tool Release

Developer launches open-source local-first workflow automation platform with Ollama integration

A developer released v0.11.0 of an open-source workflow automation platform that treats Ollama as a first-class provider, enabling local LLMs to power deterministic automation pipelines. The platform supports agent semantic memory, document RAG, visual workflow builder, conditional branching, human-in-the-loop approval nodes, and multi-agent orchestration.

2 engagement1 sourceredditMon, Jul 27, 2026, 07:03 AM

Tool Release

ISONGraph serialization format improves GraphRAG multi-hop accuracy from 40% to 80%

A developer benchmarked 10 graph serialization formats for feeding knowledge graphs to LLMs, finding that format choice alone swings multi-hop accuracy from 40% to 80% and token cost by ~70%. JSON was among the worst performers. The author built ISONGraph, a property-graph format optimized for LLM comprehension, released under MIT license in 6 languages.

5 engagement2 sourcesredditMon, Jul 27, 2026, 06:44 AM

Tool Release

Ops analyst shares prompt trick to force ChatGPT to label number sources as DATA, DERIVED, or ESTIMATE

An operations analyst posted a copy-paste prompt line that forces ChatGPT to tag every number, date, or named figure inline with its source: [DATA] if from provided data, [DERIVED] if calculated (with calculation shown), or [ESTIMATE] if from general knowledge. The trick aims to prevent users from trusting fabricated figures that appear credible.

4 engagement1 sourceredditMon, Jul 27, 2026, 06:40 AM

RSS

Tool Release

Building a coding agent harness step by step

A member-only article walks through constructing a coding agent harness, covering the agent loop, planning, subagents, sandboxing, memory, and checkpointing. It contrasts the typical failure of custom agents with the success of Claude Code on the same tasks.

0 engagement1 sourcerssMon, Jul 27, 2026, 05:01 AM

Hardware Release

AMD hits 46% server CPU revenue share with new Venice chips, shifting AI hardware focus to CPU orchestration

AMD's new Venice CPUs have propelled the company to a 46% revenue share in x86 servers, up from near zero in 2017. This shift highlights the growing importance of CPU orchestration for agentic AI workflows, which can consume up to 1,000x more tokens than basic chatbot queries.

1 engagement1 sourceredditMon, Jul 27, 2026, 04:24 AM

Tool Release

ASL V6: Open-source AST red-teaming engine for Python AI agents released

ASL V6, an open-source abstract syntax tree (AST) red-teaming engine for Python AI agents, has been released on GitHub. The tool allows developers to automatically probe AI agents for vulnerabilities by manipulating their code's AST, helping identify security weaknesses before deployment.

1 engagement1 sourcehackernewsMon, Jul 27, 2026, 03:50 AM

Tool Release

Aegis Router launches: API router for agentic IDEs with measured-energy billing

Aegis Router launched as an API router designed for coding agents, featuring measured-energy billing and cache-aware routing. It offers an OpenAI-compatible interface and aims to reduce costs for long-running agentic IDE sessions by avoiding reprocessing of repeated context.

1 engagement1 sourceredditMon, Jul 27, 2026, 03:40 AM

Model Release

Community tests Kat Coder 2.5, a Qwen-derived coding model, generating a full Star Fox-like game in one HTML file

A Reddit user tested Kat Coder 2.5, a model derived from Qwen 3.6 35B A3B, by prompting it to create a Star Fox-inspired spaceship game with vanilla Three.js. The model produced a single HTML file containing a fully playable game with ship controls, multiple enemy types, bosses, level progression, HUD, and visual effects. The user ran the model at Q4_K_M quantization and noted it consistently outperformed its base model in their tests.

37 engagement1 sourceredditMon, Jul 27, 2026, 03:26 AM

RSS

Agent Release

OMP: Open-Source 55k-Line Rust Coding Agent with Built-in Debugging and Subagents

OMP is an open-source coding agent written in 55,000 lines of Rust, featuring built-in debugging, LSP integration, browser automation, memory, and subagents. It aims to surpass AI IDEs by offering a complete, self-contained development environment.

0 engagement1 sourcerssMon, Jul 27, 2026, 03:18 AM

RSS

Tool Release

Kitewright releases 7 MB browser automation binary for AI agents

Kitewright is a new 7 MB binary that gives LLM agents a real browser for navigation, screenshots, extraction, and PDF generation without the overhead of Node.js and Playwright. It is written in a compiled language for minimal footprint.

0 engagement1 sourcerssMon, Jul 27, 2026, 03:09 AM

Tool Release

Static MCP lets developers publish MCP tools as static files with sandboxed execution

A new open-source project called Static MCP allows developers to publish Model Context Protocol (MCP) tools as static files, which are then executed in a sandboxed environment. This approach simplifies deployment and enhances security by eliminating the need for a running server. The project was shared on Hacker News on July 27, 2026.

1 engagement1 sourcehackernewsMon, Jul 27, 2026, 03:08 AM

RSS

Model Release

AntLing releases Ling-3.0-flash, a 124B hybrid-linear MoE model for token efficiency

AntLing announced Ling-3.0-flash, a 124B-parameter hybrid-linear mixture-of-experts model designed to maximize useful work per token. The model targets production agent loops where latency, cache misses, and energy costs accumulate across many turns. It emphasizes token efficiency and sustainable intelligence for real-world agentic workloads.

0 engagement1 sourcerssMon, Jul 27, 2026, 02:31 AM

RSS

Tool Rel

[truncated for AI cost control]