Search AI News
Find articles across all categories and topics
316 results for "open source"
Open Source Project of the Day (Part 30): banana-slides - Native AI PPT Generation App Based on nano banana pro
Introduction "Vibe your PPT like vibing code." This is Part 30 of the "Open Source Project of the Day" series. Today we explore banana-slides ( GitHub ), open-sourced by Anionex . Have you ever found yourself the night before a presentation with a blank slide deck — full of brilliant ideas, but completely drained by the drudgery of layouts and design? Traditional AI PPT tools may be "fast," but they're often locked into preset templates, offer little freedom, and produce homogeneous results. banana-slides is built on Google's nano banana pro image generation model, delivering a "native Vibe PPT" experience: three creation paths — one sentence , outline , and page description — upload any template or materials, intelligently parse PDF/Docx/MD files, use natural language voice editing on spe
Hermes agent might be the best open source agent for local models right now
been running hermes agent by nous research for a bit now and the local model support is genuinely better than anything else ive tried in this space the thing that sold me: it has per-model tool call parsers built in, so it actually handles tool calling properly on 30B class models where openclaw and most other frameworks just fall apart. multiple people on here have confirmed its way less token hungry too the self improving skills thing is real but honcho (the learning engine) is off by default which confused me for like 2 days before I figured it out.. once you enable it in config.yaml the difference is noticeable within a few sessions some stuff worth knowing: one command install, handles python and node and everything. supports ollama, vllm, sglang out of the box. six terminal backends
Show HN: Ray – an open-source AI financial advisor that runs in your terminal
I've been using this daily for 4 months and figured others might find it useful. This is my first open source project so would love any feedback. Ray connects to your bank via Plaid, stores everything in an encrypted local SQLite database, and lets you ask questions about your finances in natural language. No cloud, no account, your data is stored on your machine. Before anything reaches the LLM, all PII is stripped — your name, companies, transaction details are redacted and replaced with tokens, then rehydrated locally in the response. The AI never sees who you are. Comments URL: https://news.ycombinator.com/item?id=47644133 Points: 6 # Comments: 2
Running OpenClaw with Gemma 4 TurboQuant on MacAir 16GB
Hi guys, We’ve implemented a one-click app for OpenClaw with Local Models built in. It includes TurboQuant caching, a large context window, and proper tool calling. It runs on mid-range devices. Free and Open source. The biggest challenge was enabling a local agentic model to run on average hardware like a Mac Mini or MacBook Air. Small models work well on these devices, but agents require more sophisticated models like QWEN or GLM. OpenClaw adds a large context to each request, which caused the MacBook Air to struggle with processing. This became possible with TurboQuant cache compression, even on 16gb memory. We found llama.cpp TurboQuant implementation by Tom Turney. However, it didn’t work properly with agentic tool calling in many cases with QWEN, so we had to patch it. Even then, the
I Can't Write Code. But I Built a 100,000-Line Terminal IDE on My Phone.
I can't write code. I'm not an engineer. I've never written a line of TypeScript. I have no formal training in computer science. But I built a 100,000-line terminal IDE — by talking to AI. Every architectural decision is mine. The code is not. It was created through conversation with Claude Code, running inside Termux on a Samsung Galaxy Z Fold6. No desktop. No laptop. Just a foldable phone and an AI that can execute commands. Today I'm releasing it as open source. GitHub: github.com/RYOITABASHI/Shelly The Problem You're running Claude Code in the terminal. It throws an error. You copy it. You switch to ChatGPT. You paste. You ask "what went wrong?" You copy the fix. You switch back. You paste. You run it. Seven steps. Every single time. The terminal and the chat live in different worlds.
We absolutely need Qwen3.6-397B-A17B to be open source
The benchmarks may not show it but it's a substantial improvement over 3.5 for real world tasks. This model is performing better than GLM-5.1 and Kimi-k2.5 for me, and the biggest area of improvement has been reliability. It feels as reliable as claude in getting shit done end to end and not mess up half way and waste hours. This is the first OS model that has actually felt like I can compare it to Claude Sonnet. We have been comparing OS models with claude sonnet and opus left and right months now, they do show that they are close in benchmarks but fall apart in the real world, the models that are claimed to be close to opus haven't even been able to achieve Sonnet level quality in my real world usage. This is the first model I can confidently say very closely matches Sonnet. And before s
[P] Implemented ACT-R cognitive decay and hyperdimensional computing for AI agent memory (open source)
Built a memory server for AI agents (MCP protocol) and implemented two cognitive science techniques in v7.5 I wanted to share. ACT-R Cognitive Decay Memory nodes fade using the base-level activation formula: B_i = ln(Sum t_j -d ) Old, rarely-accessed memories lose salience. Frequently-accessed ones stay vivid. This keeps agent context clean without manual pruning - only "warm" memories surface at retrieval time. Hyperdimensional Computing (HDC) Routing Agent state is encoded as XOR of three 768-dim binary hypervectors: state x role x action. Routing uses Hamming distance rather than cosine similarity - works surprisingly well for sparse, structured agent state. Background Edge Synthesis A background process autonomously discovers and links semantically similar memory nodes. The graph selfo
Google open sources Gemma 4 AI models that outperform models 20x their size | The models work with near-zero latency | Inshorts - inshorts.com
Google open sources Gemma 4 AI models that outperform models 20x their size | The models work with near-zero latency | Inshorts inshorts.com
TigerFS Mounts PostgreSQL Databases as a Filesystem for Developers and AI Agents
TigerFS is a new experimental filesystem that mounts a database as a directory and stores files directly in PostgreSQL. The open source project exposes database data through a standard filesystem interface, allowing developers and AI agents to interact with it using common Unix tools such as ls, cat, find, and grep, rather than via APIs or SDKs. By Renato Losio
Monarch v3: 78% Faster LLM Inference with NES-Inspired KV Paging
TL;DR: We implemented NES-inspired memory paging for transformers. On a 1.1B parameter model, inference is now 78% faster (17.01 → 30.42 tok/sec) with nearly zero VRAM overhead. The algorithm is open source, fully benchmarked, and ready to use. The Problem KV cache grows linearly with sequence length. By 4K tokens, most of it sits unused—recent tokens matter far more than old ones, yet we keep everything in VRAM at full precision. Standard approaches (quantization, pruning, distillation) are invasive. We wanted something simpler: just move the old stuff out of the way. The Solution: NES-Inspired Paging Think of it like a Game Boy's memory banking system. The cache is split into a hot region (recent tokens, full precision) and a cold region (older tokens, compressed). As new tokens arrive,
Quoting Daniel Stenberg
The challenge with AI in open source security has transitioned from an AI slop tsunami into more of a ... plain security report tsunami. Less slop but lots of reports. Many of them really good. I'm spending hours per day on this now. It's intense. Daniel Stenberg , lead developer of cURL Tags: daniel-stenberg , security , curl , generative-ai , ai , llms , ai-security-research
Quoting Greg Kroah-Hartman
Months ago, we were getting what we called 'AI slop,' AI-generated security reports that were obviously wrong or low quality. It was kind of funny. It didn't really worry us. Something happened a month ago, and the world switched. Now we have real reports. All open source projects have real reports that are made with AI, but they're good, and they're real. Greg Kroah-Hartman , Linux kernel maintainer ( bio ), in conversation with Steven J. Vaughan-Nichols Tags: security , linux , generative-ai , ai , llms , ai-security-research