Anvira Runtime release notes

• The AI serves more than one request at a time. llama-server now runs several slots over one shared context pool: 4 on a GPU with a 16k context, 2 on the processor (whose context grows to 8k for this: a few hundred MB of memory for a small model), 1 when the context is smaller. Chat, Ask AI and ghost text no longer wait behind an exam or lesson being written, and every slot can still use the whole context, so a long prompt never gets less room than before.
• New setting models.parallel (default: automatic) to choose the number of slots yourself: anvira config set models.parallel 2.

Anvira Runtime 1.0.1: better answers from small models.

• llama.cpp's own settings reach the model. Apps can now pass llama.cpp sampling options with a chat request: repeat_penalty, repeat_last_n, the DRY sampler (dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n), top_k, min_p, typical_p, chat_template_kwargs (for example {"enable_thinking": false}), json_schema and cache_prompt. They go to models running on this PC only; requests to a cloud provider never carry them, since OpenAI-style APIs reject fields they do not know.
• Why it matters: tiny models (Qwen2.5 0.5B, Qwen3 0.6B) loop, and Qwen3 spends most of every answer thinking privately. Anvira Notes 1.2.3 uses these settings to answer directly (Qwen3 0.6B: about 0.7 s instead of 3 s) and to stop runaway repetition.

Windows x64 self-contained package: its own Python, ORCHA, Nomi, AICL, CPU llama-server, and the Microsoft VC++ runtime (no installs needed). The -cuda zip adds NVIDIA GPU acceleration (needs an NVIDIA driver 525+); unzip it into the same folder, or let 'anvira runtime install' fetch it. Verify with SHA256SUMS. Then run: anvira open