LIMONATA - CHOOSE AND LOAD YOUR WINDOWS MODEL - v5

1. Connect this PC with the Limonata provider installer if it is not linked yet.
2. Open https://limonata.xyz/marketplace#account and choose My listings.
3. Choose "Choose AI model" on your linked PC. Pick a model and save it.
4. Extract this ZIP, then double-click Start-Limonata-Model.cmd on that GPU PC.
5. Keep the window open. The launcher checks free memory, downloads the selected
   model, verifies every file, and tests full GPU offload plus a model response.
   It also checks automatic and forced structured calls without executing tools.
6. The account page shows progress. When ready, choose Open chat / API to test
   the selected model from your laptop. The session stays open until you close it manually.

OpenCode: this launcher accepts up to 2048 output tokens with an 8192-token
context. In the workspace API tab, create a fresh test key and copy the generated
configuration into your project's opencode.json. It selects the Limonata model
and a focused coding agent. Start a new conversation and select Limonata.
Try: Create hello.txt containing HELLO, then read it back.
Tools execute in your client, never on the GPU provider PC. Check the actual file
and tool results; a chat message saying DONE alone does not confirm a file change.
The site converts complete native Qwen tool envelopes into structured calls;
incomplete calls and malformed arguments return an error. Model reliability can
still vary by task. TOOL CALLS: PASS is a startup check, not a coding benchmark.
When updating from v1, v2, v3 or v4, stop the old launcher and end its private test first.
Verified runtime and model downloads are reused. The v1 launcher keeps its
512-token limit. Versions 2 and 3 accept 2048. Version 3 fixes the Coder 7B
tool template and constrains forced calls to the requested tool schema.
Restart with v5 for up to 5 minutes per model response, request heartbeats and
cancellation during inference. Versions 1-4 retain their 90-second request cap.
Verified model downloads are reused. The session itself remains unlimited.
Version 4 introduced unlimited sessions and Coder 14B. The template is derived from the
pinned Qwen GGUF template (Apache-2.0), embedded and verified by this launcher.
It is written only into the private run directory and removed at shutdown.

This release prepares and tests models. It does not publish an offer or enable
paid rentals. Existing payment and rental availability gates remain in place.
One connected listing is required by this initial launcher. It reuses that
connection without writing provider configuration or creating new keys.

Catalog: Qwen 2.5 0.5B (connection check), 1.5B (light chat), 7B (general chat),
and Qwen 2.5 Coder 7B / 14B (programming), in Q4_K_M GGUF format. Recommended GPU memory
is an estimate, not qualification. A real local GPU response is required.
All models are publisher artifacts pinned to a commit, exact size and SHA-256.
They use Apache-2.0; see LICENSE-Qwen.txt. Runtime: llama.cpp b11979, MIT.
Model revisions and hashes are included in model-catalog.json.

Downloads are kept privately under your Windows user's LocalAppData directory.
Complete verified files are reused. An interrupted partial file is discarded.
There is no model compilation or Python installation. The launcher uses the
verified Vulkan runtime; supported drivers and Windows runtime DLLs are needed.
Nothing is installed as an administrator or added to Windows automatic startup.
The model binds authenticated HTTP on 127.0.0.1 only. The launcher connects out
to Limonata; it opens no router port. It executes no renter commands or tools.

Stop the launcher before choosing another model. Closing it stops the local
model through a Windows Job. The model has no fixed lifetime. Press Q, close the window, or use Close session
on Limonata Chat. A temporary network interruption preserves the session and
retries the connection. The process keeps its 25 percent CPU cap, bounded log
buffers, one-process Job and catalog-specific RAM cap. GPU VRAM
is checked before loading; the launcher does not enforce a hardware GPU quota.
No prompts, answers, account names, paths or API keys enter progress reports.
Access keys and chat data on Limonata expire when the private test ends.
This is not a VM and does not qualify the PC for arbitrary tenant software.

Coder 14B: about 8.99 GB of verified weights, 8192-token context, quantized Q8
attention cache and at least 10600 MiB free GPU memory. Close games and other
GPU workloads first. It is slower than 7B. Local GPU and tool probes must pass;
12 GB of installed VRAM alone does not prove enough memory is currently free.
