New genuinely open-source AI projects, license-verified and brought ashore each week.https://www.daioda.io/en-ushttps://www.daioda.io/projects/deepseek-harness/https://www.daioda.io/projects/deepseek-harness/DeepSeek's open agent harness — a TypeScript monorepo built on the Cordis plugin framework where context management, tool use, and failure recovery are all composable plugins rather than hard-coded behavior. One `npx @deepseek-ai/dsh web` gets you a local web UI; plugin repos are discoverable under the `dsh-plugin` GitHub topic, and the docs tree ships architecture guides, an AGENTS.md, and a development handbook. Officially a developer preview with breaking changes expected. · License: MIT (open). Claude Code sits in our catalog as the proprietary tool everyone stars anyway. Harness is DeepSeek's answer, and the answer is MIT. When the fastest-growing repo in GitHub's history is also one of the cleanest-licensed, the industry's excuses get thinner.Sun, 06 Sep 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/deepseek-v4-flash-vision-exp/https://www.daioda.io/projects/deepseek-v4-flash-vision-exp/DeepSeek's first multimodal model in the V4 family: the V4-Flash architecture with native visual modules incorporated and continued training to unlock visual understanding — not a bolted-on vision tower. Ten days as an API-only experiment, then the weights landed August 31 with an unusually complete repo: tokenizer, prompt-encoding reference, and a minimal PyTorch inference implementation covering the vision encoder and aligner, DFlash attention, MoE, Hyper-Connections, and the DSpark speculative-decoding path — draft and target weights in one checkpoint. · License: MIT (open). Flash was the promise, Pro was the proof, and Vision is the habit. When a lab's third straight release in a month ships the same one-kilobyte MIT file, you stop calling it a license choice and start calling it an identity.Sun, 06 Sep 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/dots3-note-preview/https://www.daioda.io/projects/dots3-note-preview/The first open-weights release in Xiaohongshu's dots3 family — the sibling line of the lab's IMO perfect-score model. A 280B-total / 16B-active MoE that takes text, images, video, and audio in and writes text out, through a 7B MoE ViT vision encoder (1.2B active) and an 800M audio encoder, with 512K tokens of context. Built, per dots studio, for 'long-term daily tasks' — the days-long agent runs its self-evaluation loop targets — and the agentic numbers land the claim: 78.4 SWE-bench Verified and 75.1 Terminal-Bench 2.1, self-reported. · License: Apache-2.0 (open). The social-media company shipped what the model labs keep promising: frontier-adjacent omni-modal weights, a real LICENSE file, and no fine print. Watch list, resolved — the way we wish they all resolved.Sun, 06 Sep 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/glm-5-3-flash/https://www.daioda.io/projects/glm-5-3-flash/For a week, an anonymous model called 'ox-alpha' quietly became the most popular model on OpenCode and OpenRouter — running, notably, on Chinese AI chips. On August 26 Z.AI claimed it: GLM-5.3-Flash, a 320B-total / 18B-active multimodal MoE mixing sparse and linear attention, taking text and images in, with day-one support across vLLM, SGLang, KTransformers, and Unsloth. Priced at roughly a tenth of GLM-5.2, positioned squarely at agentic coding. · License: MIT (open). A frontier lab ran a blind taste test, won it, and then published the recipe under MIT. The stealth launch proved the model; the license file proves the policy.Sun, 06 Sep 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/mojo/https://www.daioda.io/projects/mojo/Modular's Python-superset systems language for AI compute — one language from NumPy-flavored high-level code down to hand-tuned GPU kernels — now fully open source. The standard library and kernels had been public since 2024; on August 19, about a week after Mojo 1.0 shipped and following Qualcomm's reported acquisition of Modular, the compiler and tooling followed. The modular monorepo now holds the language, the MAX inference stack, and the docs under one license. · License: Apache-2.0 with LLVM Exceptions (entire modular monorepo) (open). A language for AI that shipped its compiler under lock was always a contradiction. Resolving it took an acquisition and four years — but the resolution is the real thing, not a source-available consolation prize.Sun, 06 Sep 2026 12:00:00 GMTExperimental & Unusualhttps://www.daioda.io/projects/deepseek-v4-pro-0813/https://www.daioda.io/projects/deepseek-v4-pro-0813/DeepSeek's flagship MoE reaches general availability as the 0813 build: the V4-Pro Preview architecture re-post-trained for agentic work, with the DSpark speculative-decoding module attached and a companion DSpark repo alongside. The card doesn't restate parameter counts for the GA build — the Preview's ~1.6T-total / 49B-active figures are the best public reference — and the safetensors shards total roughly 893GB, so 'Pro' means exactly what it says. · License: MIT (open). Flash was the promise; Pro is the proof. When a lab MITs its flagship — not just the efficient sibling — the license change is policy, not accident.Mon, 17 Aug 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/muse-glimmer-30b/https://www.daioda.io/projects/muse-glimmer-30b/Meta's 29.6B dense causal transformer — 27.8B of language model plus a frozen 1.8B ViT-G/14 perception encoder — distilled from its closed Muse Spark and tuned for local, always-on agent workflows: coding agents, tool calls, LLM-as-a-judge. Text-plus-image in, text out, 131K context, shipped as BF16 and two 4-bit quantizations with a DFlash speculative-decoding drafter, official ExecuTorch and GGUF builds, and day-one Ollama support. · License: Apache-2.0 (open). For a decade the question was whether Meta would ever drop the Llama license. The answer arrived quietly, in an 11.4 kB file that says Apache and means it.Mon, 17 Aug 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/qwen3-8-27b/https://www.daioda.io/projects/qwen3-8-27b/The open-weights anchor of the Qwen3.8 generation: a 27B dense vision-language model mixing Gated DeltaNet and Gated Attention across 64 layers, with 262K native context extensible to 1M and image and video input alongside text. It's the consumer-hardware release of the pair announced with the much larger Qwen3.8-Max, shipped on the official Qwen org with FP8 builds and day-one GGUF support across the ecosystem. · License: Apache-2.0 (open). Same org, two repos: the coder ships a tag and a promise, this one ships the license. The file is the policy — everything else is a press release.Mon, 17 Aug 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/deepseek-v4-flash-0731/https://www.daioda.io/projects/deepseek-v4-flash-0731/DeepSeek's 284B-total / 13B-active MoE with 256 routed experts per layer, a CSA + HCA hybrid attention stack, a 1M-token context with 384K-token max output, and native FP8 weights — a 167GB download shipping with the DSpark speculative-decoding module and native Responses API support. The 0731 build is a re-post-training pass on the April V4 preview: architecture unchanged, agentic behavior overhauled. · License: MIT (open). Two batches ago we shelved DeepSeek-V3 for its license riders. V4-Flash answers the only way that counts: an MIT file in the repo root.Mon, 10 Aug 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/kroma/https://www.daioda.io/projects/kroma/Lodestones — the community fine-tuner behind the Chroma line — retrains Krea 2 into Kroma, with the Krea 2 Turbo delta merged in as a rank-512 LoRA. v0.2 abandons the v0.1 LoRA-plus-diff packaging for a single full checkpoint: load it in any recent ComfyUI with native Krea 2 support (Qwen3-VL text encoder + VAE), sample at 8–12 steps, CFG 1.0–1.5. · License: MIT (declared in the README for the fine-tune contribution only) (partial). A checkpoint is not a diff. Merge restricted weights into your MIT fine-tune and the restriction is what ships — this is why we read tags, not READMEs.Mon, 10 Aug 2026 12:00:00 GMTGenerative Mediahttps://www.daioda.io/projects/maple-preview/https://www.daioda.io/projects/maple-preview/Deepgrove's 20B-total / 1B-active mixture-of-experts built on ternary weights from the ground up: 24 layers, 256 experts with 8 active, a 3:1 SWA-512-to-global attention mix, and 131K context — a 5.31GB checkpoint that streams 218 tok/s on an M4 Mac mini through Deepgrove's mlx-lm fork, with a Triton/FlashAttention Transformers path for CUDA. · License: MIT (Deepgrove's mlx-lm fork and HF repo materials) (open). Bonsai compresses a big model into a phone. Maple asks the sharper question: what if it was never big to begin with?Mon, 10 Aug 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/bonsai-27b/https://www.daioda.io/projects/bonsai-27b/PrismML compresses Qwen3.6-27B to extremes: a ternary build ({−1, 0, +1}, 1.71 bits/weight, 5.9GB) keeping 94.6% of FP16 quality across 15 benchmarks, and a 1-bit build (1.125 bits/weight, 3.9GB) keeping 89.5% — running ~11 tok/s on an iPhone 17 Pro Max. Ships as GGUF for llama.cpp and MLX for Apple Silicon, with forked runtimes on GitHub. · License: Apache-2.0 (PrismML runtime forks and demo repo) (open). Colibri streams a 744B model off a hard drive; Bonsai shrinks a 27B into a phone. Two opposite bets, one thesis: the hardware you own is enough.Mon, 03 Aug 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/kimi-k3/https://www.daioda.io/projects/kimi-k3/Moonshot AI's 2.8T-total / 104B-active MoE (896 experts, top-16, 93 layers mixing KDA and gated MLA) with a 1M-token context, shipped natively in MXFP4 via quantization-aware training — a 1.56TB download. Multimodal in, agentic by design, with day-0 hosting across seven OpenRouter providers. · License: Kimi K3 License (bespoke; per model card, code and weights share it) (partial). Last week we wrote 'verify the actual LICENSE file before celebrating.' We did. Celebrate the engineering; read clause by clause.Mon, 03 Aug 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/laguna-s-2-1/https://www.daioda.io/projects/laguna-s-2-1/Poolside's 118B-total / 8B-active MoE coding model with a 1M-token context and switchable thinking modes — pre-trained in under nine weeks on 4,096 H200s. Ships in BF16 plus official FP8, INT4, and NVFP4 builds, an XS sibling, and community GGUFs; a free no-login demo runs at chat.poolside.ai. · License: OpenMDW-1.1 (HF repo materials) (partial). poolside spent years closed. Its first open drop uses the most open model license on the shelf — the only thing between it and 'open' is OSI's docket.Mon, 03 Aug 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/hy3/https://www.daioda.io/projects/hy3/Tencent Hunyuan's 295B-total / 21B-active MoE reasoner (192 experts, top-8, plus a 3.8B multi-token-prediction layer) with 256K context — released under plain Apache-2.0, a first for the Hunyuan line. Official 1-bit and 4-bit GGUF builds put a flagship-scale model on a single GPU via llama.cpp. · License: Apache-2.0 (open). Every Hunyuan release used to come with a map of where you couldn't use it. This one just comes with weights.Mon, 27 Jul 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/inkling/https://www.daioda.io/projects/inkling/Thinking Machines Lab's first open-weights model: a 975B-total / 41B-active sparse MoE (6 of 256 experts plus 2 shared) that takes text, images, and audio in, with a 1M-token context, trained on 45T tokens. Day-0 support landed in transformers, vLLM, SGLang, and llama.cpp, with an NVFP4 checkpoint alongside BF16. · License: Apache-2.0 (reference code; Tinker cookbook) (partial). The year's biggest debut is also its best license quiz. Read past the tag before you build on it.Mon, 27 Jul 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/leanstral-1-5/https://www.daioda.io/projects/leanstral-1-5/Mistral's open-weights proof-engineering model for Lean 4 — a 119B-total / 6.5B-active MoE with 256K context that writes and repairs formal proofs and specifications. It saturates miniF2F, solves 587 of 672 PutnamBench problems, and surfaced five previously unknown bugs in real-world software-verification work. · License: Apache-2.0 (open). The quiet counterexample to 'you can't trust model output': here the compiler is the fact-checker, and it never gets tired.Mon, 27 Jul 2026 12:00:00 GMTExperimental & Unusualhttps://www.daioda.io/projects/cosmos-3/https://www.daioda.io/projects/cosmos-3/A unified world model for physical AI: video generation, physical reasoning, and action generation in one architecture. Ships as Nano (8B reasoner + 8B generator) for efficient inference and Super (32B + 32B) for large-scale synthetic data, with a cookbook, training framework, and post-training scripts across the Cosmos ecosystem. · License: OpenMDW-1.1 (umbrella repo); Apache-2.0 (cosmos-predict/transfer repos) (partial). When the biggest chipmaker on earth picks a brand-new license for its crown-jewel models, the license is the news. This one reads better than almost any corporate model license before it.Mon, 20 Jul 2026 12:00:00 GMTRobotics & Embodied AIhttps://www.daioda.io/projects/hermes-agent/https://www.daioda.io/projects/hermes-agent/Nous Research's self-improving, model-agnostic agent runtime. After a session it evaluates what worked and distills it into reusable skill files, layered over a four-part memory system (session, persistent, skills, user modeling), six execution backends from local to Docker to Modal, and frontends spanning CLI, Telegram, Discord, Slack, WhatsApp, and Signal. · License: MIT (open). OpenClaw proved everyone wants a personal agent. Hermes is the bet that the one that remembers you wins.Mon, 20 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/qwen3-coder-next/https://www.daioda.io/projects/qwen3-coder-next/Alibaba's coding MoE: 80B total parameters with only 3B active, a hybrid Gated DeltaNet + gated attention layout across 512 experts, and 262K context natively. Scores 70.6% on SWE-Bench Verified and 44.3% on SWE-Bench Pro — territory that recently demanded 10–20x the active compute. · License: Apache-2.0 (declared; see notes) (open). Every few months the compute floor for real coding agents drops through another story. This is the current basement.Mon, 20 Jul 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/ai-job-search/https://www.daioda.io/projects/ai-job-search/A fork-and-fill framework built on Claude Code: populate a profile, and structured workflows search job portals, score fit, tailor LaTeX CVs and cover letters, prep interviews, and run skill-gap analysis. · License: MIT (open). Less interesting as a job tool than as a genre: personal agent workflows, versioned in git, forked like dotfiles.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/andrej-karpathy-skills/https://www.daioda.io/projects/andrej-karpathy-skills/A single CLAUDE.md distilling Andrej Karpathy's public observations on LLM coding failure modes into four rules — think before coding, simplicity first, surgical changes, goal-driven execution — installable as a plugin or dropped into any project. · License: MIT (open). Included less as a recommendation than as an exhibit: the moment prompt files started accumulating stars like frameworks.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/aurora/https://www.daioda.io/projects/aurora/Microsoft Research's foundation model for the Earth system, pretrained on large atmospheric datasets and fine-tuned into specialized versions for weather, air pollution, and ocean-wave forecasting. Aurora 1.5 extends it with more variables and hourly forecasts. · License: MIT (partial). AI weather models are quietly rewriting a field that ran on supercomputers for fifty years. This is one of the few you can actually pip install.Sun, 19 Jul 2026 12:00:00 GMTScience & Datahttps://www.daioda.io/projects/boltz/https://www.daioda.io/projects/boltz/A family of open biomolecular interaction models. Boltz-1 was the first fully open-source structure predictor to approach AlphaFold3 accuracy; Boltz-2 jointly predicts complex structure and binding affinity, claimed to be the first deep-learning model approaching physics-based FEP accuracy while running roughly 1000x faster. · License: MIT (open). The most consequential open release on this list for people who don't care about AI at all — they care about drugs, and this changes their cost curve.Sun, 19 Jul 2026 12:00:00 GMTScience & Datahttps://www.daioda.io/projects/bumblebee/https://www.daioda.io/projects/bumblebee/A read-only scanner for developer machines that inventories lockfiles, installed packages across npm, PyPI, Go, RubyGems, and Composer, browser and editor extensions, and MCP server configurations — so when an advisory names a compromised package, you can answer 'which machines have it?' immediately. · License: Apache-2.0 (open). The MCP boom created a new attack surface faster than anyone built defenses. This is the defense showing up.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/chatterbox/https://www.daioda.io/projects/chatterbox/A production-grade open TTS family: a 500M English model with unique emotion-exaggeration control, a 23-language multilingual variant with zero-shot voice cloning, and a 350M low-latency Turbo model with paralinguistic tags like [laugh]. Consistently benchmarks at or above closed competitors in blind listening tests. · License: MIT (open). The open-weights voice race ended faster than anyone expected. This is what winning it looks like.Sun, 19 Jul 2026 12:00:00 GMTAudio & Voicehttps://www.daioda.io/projects/codebase-memory-mcp/https://www.daioda.io/projects/codebase-memory-mcp/An MCP server that indexes a repository into a persistent knowledge graph: tree-sitter parsing across 158 languages, deeper type resolution for 12, sub-millisecond graph queries, dead-code detection, and architecture analysis — shipped as a single zero-dependency binary. · License: MIT (open). If your agent burns most of its budget re-reading files, this is the cure — and its release engineering should embarrass much bigger projects.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/colibri/https://www.daioda.io/projects/colibri/A pure-C inference engine that runs GLM-5.2 — a 744-billion-parameter mixture-of-experts model — on roughly 25GB of RAM by treating VRAM, RAM, and disk as one memory hierarchy and streaming expert weights from storage on demand. Implements MLA attention, speculative decoding, grammar-forced output, and optional CUDA/Metal. · License: Apache-2.0 (open). Nobody needs this. That's not the point. The point is that the wall between consumer and datacenter hardware is made of engineering, not physics.Sun, 19 Jul 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/comfyui/https://www.daioda.io/projects/comfyui/A node-graph GUI, API, and backend for running generative models locally. In 2026 it is the universal front end for open media: image (FLUX.2, Z-Image, Qwen-Image), video (Wan 2.2, LTX-2), audio, and 3D, composed into reusable visual workflows shareable as JSON. · License: GPL-3.0 (open). Learn the graph once and every new model becomes a weekend, not a migration.Sun, 19 Jul 2026 12:00:00 GMTGenerative Mediahttps://www.daioda.io/projects/deepseek-v3/https://www.daioda.io/projects/deepseek-v3/A 671-billion-parameter mixture-of-experts model (37B active per token) with multi-head latent attention and multi-token prediction — one of the strongest open-weight frontier-class models, distributed with paper and inference guidance. · License: MIT (partial). Historically important and technically superb — and the exact reason this site prints weights licenses next to code licenses.Sun, 19 Jul 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/exo/https://www.daioda.io/projects/exo/Auto-discovers the machines on your network, shards a model across them with tensor parallelism, and — as of the 2026 relaunch — supports RDMA over Thunderbolt 5 on Apple Silicon for near-datacenter interconnect speeds. Exposes OpenAI-, Claude-, and Ollama-compatible APIs. · License: Apache-2.0 (open). Datacenter tricks arriving in living rooms. The gap between 'my hardware' and 'their hardware' keeps shrinking.Sun, 19 Jul 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/firecrawl/https://www.daioda.io/projects/firecrawl/An engine to search, scrape, crawl, and interact with the web for AI use: converts URLs — including JavaScript-heavy pages — into clean markdown or structured JSON, handles site crawling and batch scraping, and offers an autonomous agent mode for data gathering. · License: AGPL-3.0 (core); MIT SDKs (open). Every agent needs eyes on the web; most teams stopped building their own the day this got good.Sun, 19 Jul 2026 12:00:00 GMTRetrieval & Knowledgehttps://www.daioda.io/projects/flux2/https://www.daioda.io/projects/flux2/Black Forest Labs' successor to FLUX.1, covering text-to-image plus single- and multi-reference editing in one architecture. The distilled klein 4B runs on ~8GB VRAM under Apache-2.0, while the 32B dev model sets the open-weights quality bar under a non-commercial license. · License: Apache-2.0 (partial). A family portrait of the open-weights business model: give away the small one, license the big one. Read the label on each box.Sun, 19 Jul 2026 12:00:00 GMTGenerative Mediahttps://www.daioda.io/projects/gemini-cli/https://www.daioda.io/projects/gemini-cli/Brings Gemini into the shell: query and edit large codebases, generate apps from PDFs or sketches, and automate ops tasks. Ships built-in file, shell, and web tools, MCP server support, and GitHub workflow integration. · License: Apache-2.0 (open). The cleanest open-source counterpoint to proprietary coding agents, and the easiest way to feel what agentic coding is without spending a cent.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/grok-build/https://www.daioda.io/projects/grok-build/A fullscreen, mouse-interactive terminal harness (written in Rust) that understands a codebase, edits files, runs shell commands, searches the web, and manages long-running tasks — interactively or headless. Supports the Agent Client Protocol. · License: Apache-2.0 (open). Open source as a distribution strategy rather than a community project — worth knowing the difference, and still one of the most interesting Rust TUIs shipping today.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/hallmark/https://www.daioda.io/projects/hallmark/A design skill for Claude Code, Cursor, and Codex that pushes back on generic AI output: 22 themes across four genres, four verbs (build, audit, redesign, study), a 65-gate quality checklist, and per-project memory to stop repetitive design choices. Pure HTML/CSS/JS, no build step. · License: MIT (open). The smallest project in this catalog and one of the most provocative: a bet that design judgment can be versioned.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/heartmula/https://www.daioda.io/projects/heartmula/A family of open music foundation models: a multilingual lyrics-and-tags-conditioned song generator (3B open, 7B internal), the HeartCodec 12.5Hz music codec, a lyrics transcriber, and a CLAP-style audio-text aligner. Generates full songs with vocals at roughly real-time speed. · License: Apache-2.0 (open). Watch what indie musicians do with this in a year. Open weights change who gets to experiment.Sun, 19 Jul 2026 12:00:00 GMTAudio & Voicehttps://www.daioda.io/projects/jan/https://www.daioda.io/projects/jan/A desktop ChatGPT alternative that runs entirely offline: a polished chat UI, a local model manager built on llama.cpp, assistants, MCP tool support, and an OpenAI-compatible local API server — with optional cloud model connections when you want them. · License: Apache-2.0 (open). The app to install for someone who wants ChatGPT without the account, the cloud, or the subscription.Sun, 19 Jul 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/khoj/https://www.daioda.io/projects/khoj/An AI 'second brain' that indexes your personal corpus — PDFs, Markdown, org-mode, Notion, Word docs — for semantic search and chat. Custom agents with personas and tools, scheduled automations, and clients for browser, Obsidian, Emacs, desktop, and WhatsApp, with local LLMs for zero-cloud operation. · License: AGPL-3.0 (open). The 'personal AI' most products only promise: your data, your machine, your rules — and it actually ships.Sun, 19 Jul 2026 12:00:00 GMTExperimental & Unusualhttps://www.daioda.io/projects/kokoro/https://www.daioda.io/projects/kokoro/A text-to-speech model that punches absurdly above its weight: natural speech in 9 languages with multiple voices, voice blending, and speed control — small enough to run in real time on a CPU, a phone, or a browser tab. · License: Apache-2.0 (open). Proof that 'small and finished' is a feature. Some models are infrastructure; this one is a brick you build with.Sun, 19 Jul 2026 12:00:00 GMTAudio & Voicehttps://www.daioda.io/projects/krita-ai-diffusion/https://www.daioda.io/projects/krita-ai-diffusion/A plugin that embeds generative AI into a real digital-painting workflow: selection-based inpainting, live painting where the AI reinterprets your canvas as you draw, ControlNet guidance from scribbles and poses, regional prompts via layers, and 4K upscaling — all on a local ComfyUI backend. · License: GPL-3.0 (open). A preview of how creative software absorbs these models: not as a button that makes art, but as another tool that listens.Sun, 19 Jul 2026 12:00:00 GMTGenerative Mediahttps://www.daioda.io/projects/langchain/https://www.daioda.io/projects/langchain/The original LLM application framework, now consolidated around durable agent runtimes: abstractions for tools, retrieval, memory, and multi-step agents across every major model provider, in Python and JavaScript. · License: MIT (open). Not the newest thing in the room anymore — which is exactly why it's safe to build on.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/langflow/https://www.daioda.io/projects/langflow/A visual canvas for composing LLMs, vector stores, tools, and APIs into agents and workflows, each servable as an API. Python-extensible, with MCP support on both the client and server side. · License: MIT (open). The visual builder to reach for when the license fine print on the alternatives gives you pause.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/lerobot/https://www.daioda.io/projects/lerobot/Hugging Face's end-to-end PyTorch library for real-world robotics: standardized datasets, hardware interfaces, and imitation- and reinforcement-learning policies for physical robots. Bundles vision-language-action models and a hardware-agnostic interface for teleoperation, data collection, training, and deployment. · License: Apache-2.0 (open). The Hugging Face playbook applied to atoms: shared datasets, shared weights, shared format. Robotics needed exactly this.Sun, 19 Jul 2026 12:00:00 GMTRobotics & Embodied AIhttps://www.daioda.io/projects/lightrag/https://www.daioda.io/projects/lightrag/A graph-augmented RAG framework from HKU's data-science lab: extracts entity-relationship knowledge graphs from documents and combines dual-level retrieval with vector search. Ships as a Python library plus a server with an interactive graph visualizer, supporting many storage backends and LLM providers. · License: MIT (open). The version of graph-RAG you can actually afford to run on your whole corpus — which is the only version that matters.Sun, 19 Jul 2026 12:00:00 GMTRetrieval & Knowledgehttps://www.daioda.io/projects/llamafile/https://www.daioda.io/projects/llamafile/Packages an LLM's weights and the llama.cpp engine into a single cross-platform executable using Cosmopolitan Libc. One downloaded file runs the same model on macOS, Linux, Windows, and the BSDs — no install, no runtime, no dependencies. A sibling, whisperfile, does the same for speech-to-text. · License: Apache-2.0 (open). Fifty years from now, when someone wants to run 2020s AI, this is the format that will still boot.Sun, 19 Jul 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/ltx-2/https://www.daioda.io/projects/ltx-2/A DiT-based audio-video model — the first open-weights system generating synchronized sound and picture together, up to 4K at 50fps. The 22B LTX-2.3 (with a fast distilled variant) covers text-to-video, image-to-video, interpolation, camera-control LoRAs, and ships an official LoRA trainer. · License: Apache-2.0 (partial). Sound changes everything about generated video — silence was the uncanny part. First open model to figure that out.Sun, 19 Jul 2026 12:00:00 GMTGenerative Mediahttps://www.daioda.io/projects/marimo/https://www.daioda.io/projects/marimo/A reactive Python notebook stored as pure .py files. Cells form a dataflow graph, so a change automatically re-runs everything that depends on it — no hidden state. Notebooks run as scripts, deploy as apps, query with SQL, and version cleanly in git, with an AI-native editor built in. · License: Apache-2.0 (open). Everyone complains about notebook reproducibility and then opens Jupyter anyway. This is the one worth switching for.Sun, 19 Jul 2026 12:00:00 GMTScience & Datahttps://www.daioda.io/projects/mattergen/https://www.daioda.io/projects/mattergen/Microsoft's generative diffusion model for inorganic materials across the periodic table. It proposes novel stable crystal structures and can be fine-tuned to steer generation toward property constraints — band gap, bulk modulus, magnetic density, or target chemistry. · License: MIT (open). Generative AI pointed at the periodic table instead of the timeline. The quiet, world-changing corner of the field.Sun, 19 Jul 2026 12:00:00 GMTScience & Datahttps://www.daioda.io/projects/mistral-rs/https://www.daioda.io/projects/mistral-rs/A fast, flexible inference engine in Rust that — despite the name — runs many model families across text, vision, and speech. Quantization (GGUF, GPTQ, and in-situ), LoRA and X-LoRA adapters, paged attention, Rust and Python APIs, and an OpenAI-compatible server with MCP support. · License: MIT (open). The engine you discover when Python dependency hell finally wins. Deserves several times its stars.Sun, 19 Jul 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/nanochat/https://www.daioda.io/projects/nanochat/Karpathy's minimal, hackable, full-stack LLM pipeline: tokenizer, pretraining, finetuning, evaluation, inference, and a chat UI, training a GPT-2-class conversational model on one 8xH100 node in about 90 minutes for $50-100. · License: MIT (open). The most valuable output isn't the model — it's the mental model. Everyone who trains one stops being afraid of the stack.Sun, 19 Jul 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/nao/https://www.daioda.io/projects/nao/An open framework for building and deploying analytics agents. You build a context layer over your data with the nao-core CLI, then deploy a chat interface so non-technical teammates can ask natural-language questions the agent answers against the warehouse. Warehouse-agnostic; bring your own LLM. · License: Apache-2.0 (open). The interesting bet here is architectural: fix the semantics first, and the agent stops hallucinating your revenue. Early, but pointed the right way.Sun, 19 Jul 2026 12:00:00 GMTScience & Datahttps://www.daioda.io/projects/officecli/https://www.daioda.io/projects/officecli/A from-scratch document engine in a single self-contained binary: reads, edits, and creates .docx, .xlsx, and .pptx, renders to HTML or PNG, evaluates Excel formulas, and runs template merges — no Microsoft Office installation anywhere. · License: Apache-2.0 (open). Agent-native tooling is becoming its own software category. This is what it looks like when someone rebuilds a 40-year-old product assuming the user is a machine.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/ollama/https://www.daioda.io/projects/ollama/A local LLM runner that packages weights, configuration, and a REST API behind a Docker-like pull-and-run workflow. Runs the current generation of open-weight models — Kimi, GLM, DeepSeek, gpt-oss, Qwen, Gemma — on macOS, Linux, and Windows. · License: MIT (open). If you recommend one tool to someone who has never run a model locally, it's still this one.Sun, 19 Jul 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/olmo-3/https://www.daioda.io/projects/olmo-3/Ai2's family of genuinely open models — 7B and 32B, in Base, Instruct, and Think reasoning variants — built on the OLMo-core training stack. Uniquely, Ai2 releases the entire model flow: training data, code, recipes, logs, and every intermediate checkpoint, all Apache-2.0. · License: Apache-2.0 (open). When someone asks what 'open source AI' should mean, point here and stop arguing.Sun, 19 Jul 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/omniroute/https://www.daioda.io/projects/omniroute/A self-hosted AI gateway exposing a single OpenAI-compatible endpoint across 231+ providers — around 90 with free tiers — with automatic fallback, cost/speed/quality routing, token compression, and ready-made hookups for Claude Code, Cursor, Cline, and other coding agents. · License: MIT (open). Genuinely useful infrastructure wrapped in a value proposition that works right up until the providers decide it doesn't. Route accordingly.Sun, 19 Jul 2026 12:00:00 GMTLocal & Private AIhttps://www.daioda.io/projects/openclaw/https://www.daioda.io/projects/openclaw/A self-hosted assistant that runs on your own devices, with a local gateway bridging AI models to the places you already talk: WhatsApp, Telegram, Slack, Discord, Signal, iMessage. Voice, a live canvas UI, multi-agent routing, and companion apps for macOS, iOS, and Android. · License: MIT (open). The clearest signal yet that the assistant wars will be fought partly on your own hardware. Extraordinary momentum — bring your own security discipline.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/openevolve/https://www.daioda.io/projects/openevolve/An open evolutionary coding agent in the lineage of DeepMind's AlphaEvolve: it uses LLMs as mutation operators inside a quality-diversity loop with island-based parallel populations, evolving programs against automated evaluators. It has reproduced state-of-the-art results and discovered real GPU-kernel speedups. · License: Apache-2.0 (open). AlphaEvolve made headlines and stayed closed. This is the version you can point at your own bottleneck tonight.Sun, 19 Jul 2026 12:00:00 GMTExperimental & Unusualhttps://www.daioda.io/projects/openpi/https://www.daioda.io/projects/openpi/Physical Intelligence's open release of its flow-based vision-language-action models. π0 is pretrained on 10,000+ hours of robot data; π0.5 adds open-world generalization via 'knowledge insulation'; π0-FAST is an autoregressive variant. Base models ship with DROID, ALOHA, and LIBERO checkpoints. · License: Apache-2.0 (open). When a robotics-foundation-model company open-sources its actual policy, the field moves. This is that release.Sun, 19 Jul 2026 12:00:00 GMTRobotics & Embodied AIhttps://www.daioda.io/projects/openwiki/https://www.daioda.io/projects/openwiki/A CLI from LangChain that writes and continuously maintains agent-facing documentation. Two modes: a living wiki of your repository, and a personal knowledge wiki fed by connectors like Git, Notion, Gmail, and web search. · License: MIT (open). The llms.txt idea grown into a system: docs as the interface between your codebase and its new robot readers.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/qwen-image/https://www.daioda.io/projects/qwen-image/A 20B image foundation model spanning text-to-image and precise instruction-based editing. Its signature strengths: complex text rendering — long passages, real typography, Chinese included — and identity-preserving edits. The 2.0 release added native 2K output and multi-image composition. · License: Apache-2.0 (open). Typography was the tell that gave AI images away. That tell is gone, and the model that killed it is free.Sun, 19 Jul 2026 12:00:00 GMTGenerative Mediahttps://www.daioda.io/projects/ragflow/https://www.daioda.io/projects/ragflow/An open RAG engine centered on deep document understanding — layout-aware parsing of PDFs, tables, and scans — fused with agent capabilities to build a context layer for LLMs, with citation-grounded answers. · License: Apache-2.0 (open). The unglamorous truth of RAG is that parsing beats prompting. This project is built on that truth.Sun, 19 Jul 2026 12:00:00 GMTRetrieval & Knowledgehttps://www.daioda.io/projects/rf-detr/https://www.daioda.io/projects/rf-detr/Roboflow's real-time transformer architecture for detection and instance segmentation, built on a DINOv2 backbone and reported state-of-the-art on COCO and RF100-VL. Six detection sizes from Nano to 2XLarge, optimized for fine-tuning on custom datasets. Accepted at ICLR 2026. · License: Apache-2.0 (core + Nano-Large models) (open). The detector to reach for when legal asks 'what's our YOLO exposure?' and you'd rather not have that meeting.Sun, 19 Jul 2026 12:00:00 GMTComputer Visionhttps://www.daioda.io/projects/ruview/https://www.daioda.io/projects/ruview/A WiFi-sensing platform that turns Channel State Information from cheap ESP32 nodes into presence detection, activity recognition, and claimed vital-sign monitoring — no cameras, no wearables. Integrates with Home Assistant, Apple Home, Google Home, and Alexa, with hardware starting around $8. · License: MIT (open). Delightful and unsettling in equal measure. Included because it's genuinely novel — and flagged because novelty isn't the same as validation.Sun, 19 Jul 2026 12:00:00 GMTExperimental & Unusualhttps://www.daioda.io/projects/sam3/https://www.daioda.io/projects/sam3/Meta's unified 848M-parameter foundation model for promptable segmentation in images and video: a DETR-based detector fused with the SAM 2 tracker, so one model detects, segments, and tracks from text prompts, points, boxes, or masks — with open-vocabulary concepts trained over ~270k unique terms. · License: Apache-2.0 (partial). Technically indispensable, legally 'read the license first.' Both things are true, and this site exists to say so in the same breath.Sun, 19 Jul 2026 12:00:00 GMTComputer Visionhttps://www.daioda.io/projects/smollm3/https://www.daioda.io/projects/smollm3/Hugging Face's fully open 3-billion-parameter model with a toggleable extended-thinking mode, 64k context natively (128k+ via YaRN), six languages, and tool calling — trained on 11T tokens with the complete recipe, configs, and data mix published. · License: Apache-2.0 (open). Small models are where open source wins outright — and this one ships with its homework attached.Sun, 19 Jul 2026 12:00:00 GMTModels & Weightshttps://www.daioda.io/projects/strix/https://www.daioda.io/projects/strix/Autonomous penetration-testing agents that behave like attackers: they run code against your application, validate findings with working proof-of-concept exploits rather than static analysis, cover source review and web assessment, classify against OWASP, and can propose patches. · License: Apache-2.0 (open). Red-teaming is being automated whether defenders are ready or not. Better to know what these agents can do to your app before someone else's does.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/supervision/https://www.daioda.io/projects/supervision/Roboflow's model-agnostic utility library: reusable building blocks for detection, tracking, annotation, dataset conversion, and video analytics. Connects to models from Ultralytics, Transformers, and MMDetection, and provides annotators, multi-object tracking, zone counting, and speed estimation. · License: MIT (open). Every field has one library everyone quietly depends on. In vision analytics, you're looking at it.Sun, 19 Jul 2026 12:00:00 GMTComputer Visionhttps://www.daioda.io/projects/txtai/https://www.daioda.io/projects/txtai/An all-in-one embeddings database combining vector search, graph networks, and relational storage with SQLite-grade simplicity. On top of its semantic index it layers pipelines, workflows, agents, and RAG — all runnable fully locally and exposable as a FastAPI service. · License: Apache-2.0 (open). Most projects don't need a vector database; they need this. Years of steady releases and zero drama — the good kind of boring.Sun, 19 Jul 2026 12:00:00 GMTRetrieval & Knowledgehttps://www.daioda.io/projects/vibe-trading/https://www.daioda.io/projects/vibe-trading/From HKU's data-science lab: natural-language quantitative research that turns a hypothesis into a signal engine and backtest. Ships 50+ MCP tools, an 'Alpha Zoo' of 452 prebuilt factors, 18 free data providers, connectors to 10 brokers, and delivery to 16 chat channels — with mandate gates, kill switches, and audit ledgers built in. · License: MIT (open). The most instructive part isn't the alpha factory — it's watching a serious lab design guardrails for agents that touch money.Sun, 19 Jul 2026 12:00:00 GMTAgents & Codinghttps://www.daioda.io/projects/wan22/https://www.daioda.io/projects/wan22/Alibaba's open video suite built on a mixture-of-experts diffusion architecture: 27B text-to-video and image-to-video models, a 5B dense variant with a high-compression VAE for consumer GPUs, plus speech-to-video and character-animation variants. · License: Apache-2.0 (open). When the next version goes closed, the last open one becomes infrastructure. Expect this checkpoint to be built on for years.Sun, 19 Jul 2026 12:00:00 GMTGenerative Mediahttps://www.daioda.io/projects/z-image/https://www.daioda.io/projects/z-image/An efficient 6B single-stream diffusion transformer family. Z-Image-Turbo generates photorealistic images in 8 steps with sub-second latency on datacenter GPUs and fits in 16GB of consumer VRAM; variants cover text-to-image, instruction editing, and bilingual text rendering. · License: Apache-2.0 (open). Efficiency is the quiet frontier. The most interesting race in image models is now about doing more with less.Sun, 19 Jul 2026 12:00:00 GMTGenerative Media