AIToday
Large Language ModelsAI Coding AssistantsOpen-Source AIZenn AI/MLPublished: Oct 9, 2026, 10:00 JST

Nemotron-3-Ultra tops free agent stack, Zenn author says

Nemotron-3-Ultra tops free agent stack, Zenn author says

A Zenn writeup describes a zero-cost agent environment built from free LLMs (Nemotron-3-Ultra, LongCat, Gemini 3.8 Flash, Muse Spark 1.3) plus Herdr and Obsidian's Claudian plugin, with Nemotron-3-Ultra called the most practical.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

The writeup is less a product announcement than a field report from someone running agents daily without paying for API access. The stack is layered by role: VS Code extensions such as Antigravity, GitHub Copilot and Codex handle inline work; the terminal multiplexer Herdr monitors several agents at once; and Obsidian's Claudian plugin auto-detects installed CLIs so the note vault itself becomes a working directory.

The friction the author describes is specific to free tiers rather than to the models. Autonomous agents resend large system prompts and tool schemas every turn and keep reading project files, so a few thousand tokens of ordinary chat can become hundreds of thousands. On shared free endpoints, that consumption collides with per-day and per-minute caps — the author's complaint is that work halts right when it is going well.

Among the tools mentioned, two stand out as open items. The Claudian plugin, credited to Yishen Tu at v2.3.16, supports Claude Code, Codex CLI, Grok Build, OpenCode and Pi; the author notes OpenCode installed under WSL is not auto-detected from Windows Obsidian. Pi, credited to Mario Zechner, is described as restricted to four basic tools — read, write, edit and bash — and the author says testing is still underway.

FAQ
Which free model did the author rate most highly?
Nemotron-3-Ultra, described as consistently the most practical. The author says it correctly identified a broken dependency and proposed a fix during a Hermes Agent update failure.
What problems did the author hit with free autonomous agents?
OpenRouter's :free endpoints are shared, so daily and per-minute caps are tight. The author says Cline-style agents accumulate hundreds of thousands of tokens over 10 to 20 turns, and rate limits cut work off mid-flow.
How does the author work around those free-tier limits?
By cutting tasks into one- or two-step instructions, pointing the agent at specific file paths instead of whole directories, and leaving heavy autonomous runs to IDE extensions with recovering quotas.

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleHex cuts wrong edits 7x with evals on Quick Edits