A 70-billion-parameter AI model running on a single 4GB GPU. A Slack alternative where AI agents are first-class teammates with their own cryptographic identities. A terminal-based code review tool that catches your AI agent's mistakes before they reach GitHub.
These aren't announcements from a Big Tech keynote. They're open-source projects on GitHub, built by independent developers and small teams, quietly solving problems that the major platforms haven't addressed — or haven't addressed well enough. If you're building with AI agents in 2026, here are seven projects worth your attention this week.
AirLLM: Run 70B Models on a 4GB GPU
The problem is straightforward: a 70-billion-parameter model in 16-bit precision needs roughly 140 GB of VRAM. The largest consumer GPU tops out at 24 GB. The standard answers — quantization, distillation, pruning, or buying more GPUs — each sacrifice something.
AirLLM, created by Gavin Li and now at over 29,800 GitHub stars, takes a different route. It loads transformer layers one at a time from disk. The GPU needs only the current layer in memory, not the entire model. Peak VRAM depends on the size of a single layer, not the whole model.
The result: you can run a 70B model on a 4GB GPU without quantization, distillation, or pruning. Llama 3.1's 405B model runs on 8GB. DeepSeek-V3's 671B runs on about 12GB.
The trade-off is speed. Every layer loads from disk for every generated token, so disk I/O — not GPU compute — sets the pace. This is a batch processing tool, not an interactive chat. But for developers who need to run a specific large model on fixed hardware and can afford to wait, AirLLM removes the hardware barrier entirely.
Install it with pip install airllm and use three lines of Python to start inference. It supports Llama 2, Llama 3, Qwen, DeepSeek, Mistral, Phi, and Gemma out of the box.
TencentDB Agent Memory: Shared Memory for AI Agent Teams
Every developer who has used a coding agent knows the problem: the agent forgets. It relearns your codebase every session. If five developers each teach their agent the same lesson about your build system, that's four wasted lessons.
TencentDB Agent Memory, now at over 15,200 GitHub stars, treats agent memory as a team-level infrastructure problem. It turns conversations, documentation, and code into four persistent asset types: Chat Memory (cross-session user understanding), Skills (distilled executable experience), Wiki (document structure and relationships), and CodeGraph (code call graphs and impact scope).
The key differentiator is ownership and access control. Memory assets have four visibility levels: private (only the owner can read), team (team members can read), restricted (precise ACL-based access), and agent (for targeted equipping within a team). You can assign a "Release Skill" to the Release Agent, an "Architecture Wiki" to development Agents, and CodeGraph to Coder and Reviewer — each agent gets exactly the context it needs.
This isn't a RAG plugin. It requires three services running plus two LLM configurations, and Tencent labels it Beta. But for teams running multiple coding agents against a shared codebase, it's the clearest implementation yet of agent memory as shared team infrastructure rather than a personal scratchpad.
Buzz: Jack Dorsey's Agent-Native Team Chat
When Jack Dorsey announced Buzz on July 21, 2026, he called it "a new groupchat platform for teams of people and agents." Behind that simple description sits an ambitious open-source project from Block.
Buzz combines team chat, Git hosting, and AI agents behind one identity system built on Nostr. Every message, reaction, code event, and approval is stored as a cryptographically signed event. Human employees and AI agents receive the same basic identity structure — their own key pairs, channel memberships, and audit trails.
Agents participate as members, not bots. They can search prior discussions, open repositories, submit patches, review code, run workflows, edit shared canvases, and create channels. Buzz supports Claude Code, Codex, and Goose as agent harnesses through the open Agent Client Protocol, keeping the model choice separate from the workspace.
The project hit 7,600 GitHub stars within days of launch. It's open-source under Apache 2.0, with packaged builds for macOS, Windows, and Linux. Block itself calls it "early stages" and TechCrunch says don't port your team over yet — but the architectural direction, agents as portable identity-bearing teammates, is worth paying attention to regardless of which specific app wins.
DeepSeek-Reasonix: A Cheap Terminal Coding Agent Built for Cache Stability
DeepSeek's API is cheap — V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens, with cache-hit input dropping to $0.003. But keeping costs low across long sessions requires keeping the prefix cache warm, which most generic coding agents don't bother with.
DeepSeek-Reasonix (also published as Reasonix), at roughly 2,500 GitHub stars, is engineered around this single invariant: prefix-cache stability. Every layer of the tool is tuned to maintain byte-stable cache prefixes across long sessions. The results are striking — one real user logged 435 million input tokens with a 99.82% cache hit rate on May 1, 2026, costing approximately $12 instead of the $61 the same workload would cost without caching on V4 Flash.
Reasonix is DeepSeek-only by design. It runs in the terminal, installs via npm install -g reasonix, and works on macOS, Linux, and Windows. If your coding work is "fix this auth bug" rather than "solve this PhD proof," the combination of DeepSeek's pricing and Reasonix's cache stability makes it one of the cheapest serious coding agents available.
TuiCR: Terminal Code Review with Vim Keybindings
AI coding agents have changed who writes most code in pull requests. They haven't changed who is accountable for it. The weakest point in an agent-driven workflow is the moment between "the agent says it's done" and "I open a pull request."
TuiCR, at over 2,300 GitHub stars, renders a GitHub-style continuous diff in the terminal with full vim keybindings. You can leave line-level, range, file, and summary comments directly in the TUI, then push them as a real PR review to GitHub or GitLab. You can also export structured markdown — which makes it agent-friendly, since your coding agent can consume classified, line-anchored comments and act on them deterministically.
The key workflow is tuicr -w: reviewing uncommitted working-tree changes after an agent has finished editing but before anything is committed. You catch the agent's mistakes at the cheapest possible point — your own terminal, before the push — rather than outsourcing quality control to your teammates' patience on GitHub.
TuiCR ships with a Claude Code skill that can launch it in a tmux or Zellij split pane, so the diff opens next to the agent's chat. The same SKILL.md format works with Copilot CLI as well.
Book-to-Skill: Turn Books Into Executable Agent Skills
We've all read good books whose frameworks we forget by the time we need them. Book-to-Skill, a project that has gained significant traction with multiple independent implementations, solves this by extracting a book's frameworks, cases, and operational steps into executable AI agent skills.
The concept works like this: you feed the tool a PDF, EPUB, or markdown file. It analyzes the content, extracts core mental models, and generates a SKILL.md file with YAML frontmatter that any compatible agent — Claude Code, Copilot CLI, or Hermes Agent — can load on demand.
When you face a decision, the agent automatically calls the relevant framework. Ask "should I quit my job to start a company?" and the agent walks through the WRAP decision framework from Chip and Dan Heath's "Decisive" — widening options, testing assumptions, attaining distance, preparing to be wrong — with real case studies at each step.
The output isn't a summary. It's a structured, executable skill with chapter files loaded on-demand (so they don't count against your context budget until you ask about that topic). The project supports technical books (tool manuals, procedural guides) and non-fiction (decision frameworks, response protocols), making it useful far beyond books — internal documentation, compliance standards, and design systems all qualify.
Microsoft AI for Beginners: A Free 12-Week AI Curriculum
Not every project on this list is a cutting-edge tool. Microsoft's AI for Beginners curriculum, at over 62,000 GitHub stars, is a 12-week, 24-lesson course covering AI fundamentals — neural networks, computer vision, NLP, and AI ethics — with hands-on labs in both TensorFlow and PyTorch.
The curriculum includes practical lessons, quizzes, and Jupyter notebooks you can run in Google Colab or locally. It starts with "Hello AI World" (pattern recognition) and simple neural networks built from scratch, then progresses through image classification, text sentiment analysis, reinforcement learning, and generative AI.
It's completely free under the MIT license. No registration, no email gate, no upsell. Just clone the repo and start with lesson one. For developers moving into AI agent work who need to solidify their foundational knowledge, this is one of the best structured starting points available — and it costs nothing.
Which of These Should You Actually Try?
If you have a specific large model you need to run on hardware that can't fit it, AirLLM is the answer. If your team runs multiple coding agents against a shared codebase, TencentDB Agent Memory solves the repetition problem. If you're rethinking how your team works with agents, Buzz is worth a weekend spike. If you want a cheap terminal coding agent, pair Reasonix with a DeepSeek API key. If you're shipping AI-generated code, TuiCR closes the review gap. If you read books you wish you could operationalize, Book-to-Skill converts them. And if you're new to the fundamentals, Microsoft's curriculum is the starting line.
None of these require a credit card, a vendor relationship, or a platform migration. They're open source on GitHub, and they solve problems the major platforms haven't — or haven't solved well enough for the developers who need more.
Related reading
- Free open-source AI tools on GitHub in 2026
- How divergent ideation makes AI coding agents think 2x better
- Buzz by Jack Dorsey: a free AI agent workspace from Block
- Best free AI agent tools in 2026
- DeepSeek V4 Flash 0731: free coding agent setup guide
- 10 free open-source AI tools in 2026
Every claim here is traced to a primary source, dated, and listed under Sources. Research and drafting are AI-assisted; editing, verification and publication are human decisions, and a person is accountable for what appears on this page. How we work →
Discussion
0 comments