Ziva Baby

Select Ziva Baby in the model picker to use Ziva’s lowest-cost hosted chat lane. Baby requests have no customer charge. Hosted access still follows your account’s subscription and trial rules. Where Hobby hosted allowances are available, Baby has a daily message limit shown in your account; paid plans have no Baby daily message cap. Tool calls and automatic compaction within a turn do not spend another daily message. The daily reset is at midnight UTC.

Model allocation

Compatible plugin versions allocate Baby users approximately equally across GPT-6 Luna, Nemotron 3.5 Lightning Free, and Reflection Beam Free. Cat Jam is another name for Beam, not another choice. Allocation is stable for an account, not a promise that one third of requests use each provider. Existing chats keep their assigned model when allocation weights change. If a model is explicitly retired, the next new turn can move; an ongoing turn never silently changes models after an error.

Baby’s tooltip in the model picker and composer describes automatic per-chat allocation and shared text-only limits, not the model assigned to your chat.

Older plugin versions that do not support Baby’s shared limits continue using Luna outside this allocation. Update your plugin to use the shared profile. Selecting a concrete model instead of Baby does not enter Baby allocation.

Images and limits

Baby uses a shared text-only chat profile with a 272,000-token context window and up to 32,000 output tokens, including provider-specific reasoning. Ziva reserves context headroom, so usable conversation space is smaller. Supported images are analyzed separately and the resulting description is supplied to Baby. Baby’s chat model does not receive those image pixels. Selecting Luna directly retains Luna’s own capabilities. Baby uses provider defaults for reasoning and service tier rather than model-specific tuning controls. Select a concrete model if you need to choose its supported settings.

If a conversation is too large, compact it or start a new chat. A single oversized message or oversized tool definitions may require a smaller message instead. Manual compaction is a new turn; automatic compaction stays with the turn that triggered it.

Feedback and measurements

Use the thumbs under a reply to give feedback, change your vote or remove it. Each served reply has one current vote. Copies of that reply in forks share the same vote; a newly generated reply in a fork has its own. Feedback stays attributed to the model that served the original reply, not a model you select later. Older replies without verifiable attribution remain unattributed. When you reopen a chat or reconnect while signed in to the same account, Ziva loads your saved feedback. Changing a shared reply’s vote also changes it for its fork copies, not for new replies generated in those forks.

While saved feedback loads, the thumbs are disabled and labelled Loading saved feedback. A successfully loaded reply with neither thumb selected has no vote. Feedback unavailable means Ziva could not verify your saved feedback, not that your vote was cleared. Feedback does not change your plan, charges or message allowance.

Votes are linked to your account and the original reply. Saving feedback sends the vote, reply identifiers and usage summaries, not the reply text. Loading saved feedback sends reply and model identifiers with available original-turn attribution, not the conversation or reply text. The normal chat request still follows the provider policies below.

Latency runs from Ziva receiving a model request until its first nonempty text, reasoning or generated tool arguments arrive. Metadata and empty chunks do not count. This can be earlier than the final answer you see.

Generation speed combines comparable completed requests: total output tokens divided by their combined first-to-last output time. Reasoning tokens count once. Waiting for the first output, running tools and waiting for a finish message are excluded; pauses between output chunks remain included. Missing or incompatible usage, or a stream with no measurable output interval, makes speed unavailable, not zero.

Daily measurements use the request’s start date in UTC, even if it finishes, fails or is stopped after midnight. Error rates compare failed main-chat model requests with successful and failed requests. Cancellations are separate; admission refusals, titles, monitoring and image offloads are excluded. These measurements do not change Baby’s charges or message allowance.

Daily comparisons

Ziva reviews Baby using internal daily aggregate summaries for the preceding completed UTC day. Requests belong to the day they started. A vote belongs to the day the original reply’s turn began, not the day you clicked a thumb. Each summary uses the latest vote when it is prepared; later vote changes do not rewrite an already confirmed summary.

Summaries show each model and allocation-policy revision separately, including models with no traffic. They include user and request counts, current up/down votes, response measurements and the number of measured samples. Missing measurements remain unavailable rather than zero, and cancellations stay separate from model errors. Allocation percentages describe configured weights, not a promised share of requests.

These messages contain model-level totals, not prompts, reply text, code, email addresses or account identifiers. Ziva reviews allocation weights manually; the summaries do not automatically change your assigned model.

Privacy

Baby can use different providers with different data policies. Do not assume Baby is zero-data-retention or that prompts are excluded from training. Nemotron’s free provider may retain prompts and responses for model improvement. NVIDIA’s published access terms restrict confidential information unless an applicable agreement permits it. Reflection’s free route does not advertise zero retention or an unconditional no-training guarantee. Reflection’s policy says content is not used for general model training by default, but allows flagged content and safety signals to be retained and used for safety training and evaluation. Ordinary content can be retained for up to 30 days; flagged content can be kept longer. Luna’s route has different protections, but those do not apply to all Baby chats. Avoid confidential material unless the applicable provider terms permit your use.

Provider references:

Errors

Provider refusals, overload, timeouts, and stopped requests are not retried on a different model. If a free provider reports a charge or cannot verify a completed request’s cost, that model is blocked until its pricing is explicitly verified. There is no paid replacement. An ordinary interrupted request with no completed cost report does not by itself trigger that block.

A model-binding or profile error stops before provider dispatch rather than changing a running turn’s model. Update the plugin if asked. If feedback is unavailable, check your connection and that you are signed in to the account that gave the feedback, then reopen the chat. If a vote still cannot be verified, update the plugin. Changing models does not repair an older reply’s identity. If the error persists, report it with the correlation ID and whether it followed Stop, steering, tools, compaction, reopening or a fork. Do not include private project content in a public report.