Jivam 0.2.1 is out, and it’s the release where “bring your own model” stops meaning “pick one model at setup and live with it.” You can now add Together AI as a provider, switch models on the fly from inside a chat or coding session, and — for anyone running a Together reasoning model — actually read the response instead of scrolling past a wall of raw <think> tags. Windows setup also got the same guided, no-admin treatment macOS has had for a while.
Here’s the full rundown.
Together AI, as a first-class provider
Jivam already supported Sarvam AI, Krutrim Cloud, and Groq. Together AI joins that list — endpoint, defaults, and all — so setup is the same one-screen flow you’re used to: paste your API key, pick Together AI from the list, and go.
Custom endpoints got more forgiving at the same time. If you’re pointing Jivam at your own OpenAI-compatible server, you can now paste either the base URL (https://your-host/v1) or the full chat-completions URL (https://your-host/v1/chat/completions) — Jivam normalizes it either way instead of silently failing on whichever form it didn’t expect.
A real model picker — and you can change your mind mid-conversation
Every provider Jivam supports follows the OpenAI API convention of exposing a /v1/models endpoint. Jivam now calls it. Open Settings → Models and you’ll see a searchable dropdown of every model your provider currently offers, instead of typing a model name from memory (or a stale doc).
More importantly, that same picker now lives inside Chat mode and Code mode themselves, in the settings popover next to Deep Run and Max Iterations. Mid-conversation, decide you want a different model — done. Jivam quietly re-initializes the session behind the scenes and reloads your conversation history, so nothing you’ve already discussed is lost. Saving a new default model from the main Settings screen does the same thing immediately, rather than waiting for your next new chat.
Under the hood this isn’t a live hot-swap — jiva-core’s model orchestrator doesn’t expose one — so Jivam does the next best thing: persist the new model, cleanly tear down and reinitialize the session with the exact same workspace and MCP servers it had a moment ago, then restore the conversation. From where you’re sitting, it just feels instant.
No more raw <think> tags in your chat
Reasoning models — Together AI’s included — often narrate their own thinking inline, wrapped in <think>...</think>. Most providers strip that before it reaches you. Some don’t, and previously, if yours didn’t, that entire internal monologue landed directly in your chat window in Deep Run mode.
Jivam now catches this itself: thinking content is pulled out of the response and tucked behind a collapsible Show thinking toggle, in both Chat and Code mode, for live responses and for conversations you reload later. The answer you asked for is right there; the model’s scratch work is one click away if you want it.
Vision-capable reasoning models, correctly routed
Jivam used to assume any model that could “see” an image had to be a separate multimodal model, routing every image through a caption-then-forward pipeline before your reasoning model ever saw it. That’s not true of every provider anymore — some reasoning models handle images natively. There’s now a “this model supports vision” toggle in advanced Model settings; flip it on for a model that has native vision, and Jivam sends images to it directly instead of taking the detour.
While we were in there, we also fixed Sarvam-105B’s default output-token ceiling, which had drifted from the actual limit on Sarvam’s standard plan.
Windows setup, brought up to parity with macOS
The install experience on Windows had been playing catch-up to macOS for a couple of releases. 0.2.1 closes that gap:
- Guided install, not forced automation. Windows now gets the same graphical walkthrough macOS Safari users see —
jivam --installopens Edge to a page that shows you exactly where to click (“Install this site as an app” in the address bar), and detects automatically once it’s done. No scripted clicking, no silent fallback to a lesser experience if a permission prompt gets in the way. - Truly no-admin background service. Jivam runs as a background service so it’s already up by the time you click its icon. On Windows this used to mean a Scheduled Task — which, it turns out, requires administrator rights to register a logon trigger no matter how you configure the task’s run level. Jivam now uses a small self-restarting service tucked into your own Startup folder instead: a plain per-user file, zero elevation required, same effect.
- A setup screen that gets out of your way. The first-run screen no longer makes you click through Node.js and dependency checks before it’ll let you configure a model. Those checks still run — they just quietly stay out of sight unless something’s actually broken.
- Smarter install script. Fixed a version-detection bug that could make the installer think a perfectly current Node.js install was too old, and added a genuine no-admin fallback: if Node isn’t already there, Windows gets a portable copy extracted straight into your user profile, no installer or elevation involved.
Getting it
If you’re already running Jivam, it’ll pick up 0.2.1 automatically the next time you open it. New here? One command gets you set up — see the download page for macOS, Linux, and Windows.
As always, full technical detail lives in the release notes on GitHub.