feat: add 50 open source blog posts and author avatar support

app/[slug]/page.tsx

@@ -253,7 +253,7 @@ export default async function BlogPostPage({ params }: PageProps) {
                   className="h-11 w-11 shrink-0 overflow-hidden rounded-full ring-1 ring-neutral-200 transition-colors hover:ring-neutral-400 dark:ring-neutral-800 dark:hover:ring-neutral-600"
                 >
                   <Image
-                    src="/ossium_logo.webp"
+                    src={post.avatar ?? "/ossium_logo.webp"}
                     alt={post.author}
                     width={44}
                     height={44}
@@ -263,7 +263,7 @@ export default async function BlogPostPage({ params }: PageProps) {
               ) : (
                 <div className="h-11 w-11 shrink-0 overflow-hidden rounded-full ring-1 ring-neutral-200 dark:ring-neutral-800">
                   <Image
-                    src="/ossium_logo.webp"
+                    src={post.avatar ?? "/ossium_logo.webp"}
                     alt={post.author}
                     width={44}
                     height={44}

content/posts/README.md

@@ -23,6 +23,7 @@ publishedAt: 2026-07-11
 updatedAt: 2026-07-11
 author: ossium
 authorUrl: https://ossium.in
+avatar: /logo/opencode-logo-dark.png
 featured: false
 draft: false
 image: /demo/oss_landingpage.webp
@@ -39,6 +40,10 @@ answerSummary: "Only for type=question - short answer for FAQ rich results."
 | `howto` | Step-by-step workflows |
 | `question` | FAQ / “What is…” pages (adds FAQ JSON-LD when `answerSummary` is set) |
 
+## Author avatars
+
+- `avatar` is optional. If omitted, the post page shows the ossium logo as the author avatar.
+
 ## Tips for ranking
 
 - One primary keyword intent per post  

content/posts/best-new-open-source-launches-of-2026.md

@@ -0,0 +1,45 @@
+---
+title: "Best New Open Source Launches of 2026"
+description: "An honest roundup of open source projects that gained serious traction in 2026, grouped by AI coding tools, local models, self-hosted apps, databases, and dev tools."
+type: article
+category: tools
+tags: [new open source projects, open source launches 2026, ai coding tools, local models, self hosted apps]
+keywords: [best new open source launches 2026, new open source projects 2026, open source ai tools 2026, self hosted apps 2026, open source databases 2026]
+publishedAt: 2026-08-08
+updatedAt: 2026-08-08
+author: OpenCode
+avatar: /logo/opencode-logo-dark.png
+featured: false
+---
+
+Every year produces a handful of open source projects that go from a repo link on Hacker News to something people actually build on. This is a curated look at the ones that gained real traction in 2026, grouped by category. One honest caveat up front: I am not going to pin an exact launch date on every project. Some of these shipped years ago and hit their stride in 2026, and for those I say they gained traction recently rather than invent a release date.
+
+## AI coding tools
+
+The most visible launches this year were coding agents. [Cline](https://github.com/cline/cline) matured from a VS Code extension into a general agent with tool permissions, and its model-agnostic design means it runs on local models just as easily as on paid APIs. [OpenHands](https://github.com/All-Hands-AI/OpenHands) pushed autonomous, multi-file development further, with an agent framework that people are embedding into their own pipelines rather than just using as a chat tool.
+
+[Aider](https://github.com/Aider-AI/aider) did not launch in 2026, but it deserves a mention because it is still the reference implementation for how a terminal agent should handle git diffs. And [Qwen Code](https://github.com/QwenLM/qwen-code) shows how far open weights have come: a capable coding model and agent released as weights you can actually run yourself. There is more depth on the local side in [our coding LLM post](/what-is-the-best-local-llm-for-coding-in-2026).
+
+## Local models and open weights
+
+The model space in 2026 is not about one flagship release. It is about small, capable models that run on consumer hardware. The Qwen family stayed dominant for size-to-quality ratio, DeepSeek kept pushing reasoning models as open weights, and Meta kept iterating on Llama. One genuinely open (not just open weights) release worth watching is [OLMo](https://github.com/allenai/OLMo) from the Allen Institute for AI, because it publishes the training data and recipes along with the weights, which is still rare.
+
+Be precise about the terminology here: most of these are open weights, not open source in the OSI sense. The weights are downloadable and usable, but training data and often the training code are not public. That distinction matters and is worth a read in [how open source licensing works](/how-open-source-licensing-works).
+
+## Self-hosted apps
+
+Self-hosting had a strong 2026. [Immich](https://github.com/immich-app/immich) became the default Google Photos alternative for people who run their own servers, with steady releases improving search and face grouping. [n8n](https://github.com/n8n-io/n8n) went from a niche automation tool to a genuine Zapier competitor, helped by its fair-code license and a huge library of integrations. [PocketBase](https://github.com/pocketbase/pocketbase) keeps growing because one binary gives you auth, a database, and file storage, which is a remarkably pleasant starting point for side projects.
+
+Open WebUI continued to be the front end of choice for local LLM setups. If you want a broader list, the [31 new open source tools](/31-new-open-source-tools) post covers a wider sweep.
+
+## Databases
+
+DuckDB did not launch this year, but it is impossible to leave out of any 2026 list because it became the default tool for analytical queries on local files, to the point that data folks reach for it before CSV libraries. [SurrealDB](https://github.com/surrealdb/surrealdb) kept improving its multi-model approach, and the vector database field stayed crowded but healthy, with [Qdrant](https://github.com/qdrant/qdrant) and [LanceDB](https://github.com/lancedb/lancedb) both shipping meaningful updates. On the dataframe side, [Polars](https://github.com/pola-rs/polars) has been eating pandas' lunch for speed and stays one of the fastest-moving data projects around.
+
+## Developer tools
+
+In dev tooling, the Astral projects were the story of the year. [uv](https://github.com/astral-sh/uv) effectively replaced the entire Python tooling workflow for a huge share of developers, and [Ruff](https://github.com/astral-sh/ruff) did the same for linting and formatting. [Bun](https://github.com/oven-sh/bun) kept pushing the all-in-one JavaScript runtime, and [Biome](https://github.com/biomejs/biome) established itself as the fast replacement for the aging prettier-and-eslint combo. [oxc](https://github.com/oxc-project/oxc) is the one to watch underneath, since its Rust-based compiler and linter components are increasingly used as building blocks by the tools above.
+
+## Reading the list
+
+None of this is a guarantee that any project will stay on top. Software trends are short, and 2026 has already seen forks and rewrites overtake once-dominant tools. Judge these on your own machine, check the license, and pick what fits your workflow rather than what has the most stars this week.

content/posts/best-open-source-ai-ml-frameworks.md

@@ -0,0 +1,45 @@
+---
+title: "Best Open Source AI/ML Frameworks"
+description: "PyTorch, TensorFlow, scikit-learn, JAX, Hugging Face Transformers, LangChain, Ollama and vLLM compared honestly, including the difference between open weights and OSI open source licenses."
+type: article
+category: tools
+tags: [best open source ai frameworks, open source machine learning, pytorch, transformers, ollama, vllm, open weights]
+keywords: [best open source ai ml frameworks, open source machine learning frameworks, pytorch vs tensorflow, ollama vs vllm, open weights vs open source]
+publishedAt: 2026-08-08
+updatedAt: 2026-08-08
+author: OpenCode
+avatar: /logo/opencode-logo-dark.png
+featured: false
+---
+
+Licensing matters more in AI than anywhere else, because "open source AI" now covers two very different things: code you can legally ship and modify, and model weights that come with a document telling you how you may use them. The two are frequently confused, and the confusion has real cost when you go to production.
+
+## Deep learning frameworks
+
+PyTorch is the default for research and for most production deep learning. BSD-3 licensed, a Python-first design that beats anything else for iterating on a model, and the ecosystem around it, Hugging Face, Lightning, distributed training tooling, is the largest in the field. If you are starting today, start here.
+
+TensorFlow remains a serious production choice, Apache 2.0, especially where you already rely on its serving and deployment stack, TFX, TensorFlow Serving, and Keras for prototyping. Its research mindshare has shrunk, but it is not a legacy product. It is the right call when your team or your infrastructure is already committed to it.
+
+## Classic machine learning
+
+scikit-learn, BSD-3, is not a deep learning tool and does not want to be one. For structured data, tabular problems, regression, classification, clustering, feature engineering, it is still the first thing to reach for and often the last thing you need. A huge share of real business ML is a well-built scikit-learn pipeline, not a neural network.
+
+JAX is the research alternative, Apache 2.0, built around XLA compilation and functional transforms. It dominates the bleeding edge of new model architectures and runs happily on TPUs. Its audience is people doing novel research, and its cost is that the API is less forgiving than PyTorch. Choose it for research ambition, not for shipping a recommendation model.
+
+## The model layer
+
+Hugging Face Transformers, Apache 2.0, is less a framework and more the interface the entire open model ecosystem standardised on. One API, thousands of pretrained models, and the default way to load and fine-tune open weights. Almost every open model ships a Transformers implementation. If you work with open models at all, you will use it.
+
+LangChain, MIT, sits on top of models and chains together retrieval, prompts, tools, and agents for LLM applications. It is the most popular orchestration library and its API churn is a documented problem: examples from last year break this year. Use it where its integrations genuinely save you work, and do not assume the abstractions will survive a major version.
+
+## Running models locally
+
+Ollama, MIT, is the fastest way to run open models on a laptop or a single GPU. It wraps model runtimes like llama.cpp behind a simple CLI and an OpenAI-compatible API, so a 7B model that answers your coding questions is a one-command install. It is not built for high concurrency or careful throughput tuning.
+
+vLLM, Apache 2.0, is that production answer: a high-throughput inference server for LLMs, with continuous batching and PagedAttention, designed to serve many concurrent requests efficiently on GPUs. If you are putting an open model behind an API that real users hit, vLLM is the standard starting point. It wants a proper GPU deployment and its own operational care.
+
+## Open weights versus open source
+
+This is the accuracy point. Model weights like Llama, Qwen, and DeepSeek are released under their own licenses, for example the Llama Community License, which are not OSI-approved open source licenses. They typically restrict commercial use above certain scale, forbid building competing models, or require special terms for big companies. The weights are open in the sense that you can download them. They are not open source in the OSI sense, and calling them that is how companies end up with a surprise legal conversation after launch.
+
+The frameworks and libraries above are genuinely open source under permissive licenses. The models you run on them often are not. [Read the license](/how-open-source-licensing-works) on both before you build a product. The framework choice is a skill decision. The weights choice is a legal one, and it is the one most teams skip.

content/posts/best-open-source-alternatives-to-popular-paid-software.md

@@ -0,0 +1,38 @@
+---
+title: "Best Open Source Alternatives to Popular Paid Software (Photoshop, Office, Slack, etc.)"
+description: "Working open source replacements for Photoshop, Microsoft Office, Slack, Adobe Premiere, Jira, Zoom, Notion, and SurveyMonkey, with an honest look at what each swap costs you."
+type: article
+category: tools
+tags: [open-source-alternatives, free-software, gimp, libreoffice, mattermost]
+keywords: [open source alternatives to paid software, gimp vs photoshop, libreoffice vs microsoft office, open source slack alternative, kdenlive vs premiere pro]
+publishedAt: 2026-08-08
+updatedAt: 2026-08-08
+author: OpenCode
+avatar: /logo/opencode-logo-dark.png
+featured: false
+---
+
+There is a working open source replacement for almost every paid tool on your laptop. The hard part is not finding one, it is understanding what the swap costs. An open source tool is rarely a drop-in clone. It usually trades polish, ecosystem, or setup time for price and control. Here is the honest version of the common swaps, with a table for quick comparison.
+
+| Paid tool | Open source alternative | What you give up | What you gain |
+|---|---|---|---|
+| Adobe Photoshop | GIMP, Krita | Adobe's polish, the camera raw pipeline, the huge tutorial ecosystem | No subscription, full control of the source; Krita is genuinely excellent for illustration |
+| Microsoft Office | LibreOffice, OnlyOffice | Pixel-perfect fidelity on complex documents, deep cloud collaboration | Real file format compatibility, no license fees; OnlyOffice self-hosts with collaborative editing |
+| Slack | Mattermost, Rocket.Chat, Zulip | The app store, hosted reliability, mobile polish | Data on your own servers, unlimited message history, Zulip's threaded model |
+| Adobe Premiere | Kdenlive, Shotcut | Some pro finishing features, Premiere's plugin ecosystem | No Creative Cloud fee; DaVinci Resolve is free but proprietary, not open source |
+| Jira | Plane, Leantime | The plugin ecosystem, enterprise reporting | A cleaner UI, no per-user pricing, modern and fast |
+| Zoom | Jitsi Meet, BigBlueButton | Video quality at extreme scale, some host controls | Meetings on your own infrastructure, no time limits, encryption you control |
+| Notion | AppFlowy | Polished sync, a giant template marketplace | Local-first data, offline, no lock-in |
+| SurveyMonkey | Formbricks | Enterprise reports, a huge question library | Self-hosted forms, unlimited responses, data stays under your control |
+
+A few of these swaps deserve more than a row in a table.
+
+Design and media is where expectations go wrong. GIMP carries a reputation for its old interface, but for print and web work it does the job, and recent versions have improved. If you paint or draw, Krita is better than GIMP for illustration and holds its own against paid apps. For video, decide what you actually need. DaVinci Resolve is excellent and free, but it is not open source; the code stays closed and the studio tier is paid. If open source is a hard requirement, Kdenlive and Shotcut cover cutting, color, and export for most creators. What you lose is Resolve's node-based color grading, the reason most people use it.
+
+Office is a matter of tolerance. LibreOffice has been around long enough to read and write docx and xlsx without drama in most cases. OnlyOffice adds a web interface built for teams, and the community edition self-hosts cleanly. Both give up the last few percent of fidelity on complex documents, dense tables, and heavily formatted client deliverables. For internal documents nobody notices. Test on real files before you promise anything to a client.
+
+Chat is where the data argument does the real work. Slack hides message history behind a paywall and keeps your data on its servers. Mattermost is the closest clone, channels, threads, integrations, and it runs fine on a small server. Rocket.Chat is heavier but has a long track record, including in Indian government projects. Zulip organizes everything into topics rather than a scrolling wall, which sounds like a gimmick until your team actually lives in it.
+
+Jira is the swap that surprises people. Plane is fast and modern, closer to Linear in feel than Jira, which makes it popular with small teams that never liked Jira. Leantime is lighter still. What both lack is the ecosystem: thousands of plugins and enterprise reporting. A ten-person startup will rarely miss it; a fifty-person org under a strict audit might.
+
+The rule for switching: if the tool holds your data and you use it daily, migrate carefully and keep a test period. If you open it occasionally, the switch is low risk. And read the license before you commit a workflow to any tool, several "free" products are closed-source freeware rather than open source. My guide to how open source licensing works explains how to tell the difference.