Local AI plans · unlimited local requests ↗

New network AI: pay by token with separate Etheraio AI Credits ↗. New network pricing is pay as you go; the subscription controls below are retained only for the earlier, inactive service.

Your assistant.
Your research.

A personal Etheraio AI plan for everyday assistance and medicine research support, with model updates evaluated before release.

£4.99 / month

In development · subscriptions are not open yet. No payment is accepted.

The monthly price is set. Proposed limits are 40 messages in a rolling 3-hour window and 1,000 messages per billing month, with one request running at a time. Responses are bounded to 512 output tokens and a 4,096-token context. These draft limits need capacity and cost validation before subscriptions open. No automatic extra charges apply. Paid access will require working node inference, account usage tracking, cancellation and confirmed subscription payments.

Medicine research support first.

The mine can train a separate Qwen3-derived research adapter from checked public examples and evaluate it against held-out answers. The latest evaluated candidate failed its required answer checks and remains experimental. The public assistant uses a separately pinned Qwen3.5 base model when an authorised Etheraio host is online; it does not claim to be that trained derivative.

The personal assistant is still being connected. It is not available for treatment recommendations or clinical decisions. Descriptor calculations and generated text do not establish that a medicine is safe or effective.

Save your interest.

This saves an interest record in your Etheraio account. It does not create a subscription or authorise a charge.

No active allowance. Remaining usage and reset times appear after a verified subscription is active.

No subscription is active.

No cheapest-service guarantee is made. Features and usage limits vary between AI services. Training and sharing of private data require separate consent; subscribing will not start mining.

Your shared AI grid

Invite people to use your grid or contribute their computers. Consenting hosts process independent text and code requests. Prompts are visible to the invited host processing them; keep secrets and personal information out. Grid chats are excluded from training. More hosts increase capacity, not pooled VRAM or guaranteed model quality.

Loading your AI grids…

Ask Etheraio AI.

Available only with an active verified subscription and an online Etheraio-operated host. Your prompts are not shared with volunteer miners or used for training. Requests are removed after expiry; delivered responses expire from the service after one hour.

Checking service…

Relevant published numerical references are retrieved from Etheraio when available. Other model statements are not automatically source-verified; the service does not browse arbitrary websites or execute commands.

This assistant provides general information and research explanations. It cannot approve medicines or make clinical decisions.

One mine, sized for your device.

Hardware detection sizes reference segments and prompt batches. Compatible installed models can place layers in GPU memory and system RAM, or run on CPU. Source 8K asset detail is retained for supported tiled rendering; lower memory can mean longer runs. The complete weights and active stage must still fit admitted memory and disk budgets. Cross-device model-layer execution is not implemented.

Qwen3/3.5, Qwen coding, Qwen-derived DeepSeek R1 and OLMo model families are routed only when installed and suitable for the task. Image, mesh and speech families have separate runtime requirements. Each derivative retains its source licence and its own evaluation; names alone do not establish an improvement.

Learn from checked research.

Public numerical outcomes can prepare a Qwen3 training dataset. Incorrect attempts are paired with the recomputed answer. Private chemistry, personal details and unsupported clinical claims stay out.

Checking training evidence…

Model-weight training has executed on the prepared research profile. Candidates stay separate until their checks pass. Eligible public training can accumulate verified weight only after the authorised evaluator loads the actual adapter and verifies fresh and held-out answers; failed or local-only training earns zero. The installed working model is preserved.