Clutcha

v1.5 — a new engine and a refreshed reading experience ENGINE UPGRADE • Upgraded to the latest MLX inference engine — fixes around long-conversation caching, multi-turn memory drift and tool-call arguments • Models now load only when you start chatting, so nothing is held in memory at launch • A hard KV cache cap keeps long conversations and auto-compaction clear of the memory ceiling READING • Full LaTeX rendering — fractions, roots, matrices, cases and boxed answers now display correctly • Fixed formulas that failed to parse, and ones that vanished entirely when parsing failed (worst case now falls back to the raw text) • Markdown rendering: headings, lists, tables, quotes, bold/italic, links • Code blocks and inline code render separately; fenced code no longer mistaken for math • New "General Rendering" section to control LaTeX and Markdown rendering independently (turn Markdown off to copy raw text); charts and diagrams are enabled per prompt ANSWER LENGTH • Higher output limits: 0.6B / 1.7B from 768 → 1536, 4B models from 2048 → 3072 • When a reply hits the limit you now get a clear notice and a one-tap "Continue" WEB SEARCH (NEW, OFF BY DEFAULT) • New Web Search tool (used via /tools) • Two sources: Tavily (bring your own API key, free tier available) or Wikipedia • Results are collapsed by default, with tappable source links; the network is used only when you turn this on and run the tool COMMANDS & PANELS • All commands now open a panel you can tap through: /help /status /model /set /prompt /language /tools /kb /game /downloader • No arguments opens the panel; arguments still work as a command (e.g. /set top_p 0.8) • Settings come in three scopes: sampling (temperature / top P) follows the prompt (remembered per model × prompt); thinking, context, output, memory profile and KV reuse are remembered per model; rendering, auto-compact and network are global • "Reset to default" now resets only the current model FIXES • /language "Auto" now really follows your input language (the first reply used to come back in English) • Fixed the first exchange not appearing after starting a new session — it needed one more message before showing • Fixed "Reset to default" being capped at a 4096 context budget • Fixed thinking mode silently raising the output limit — what you set is now what you get • Qwen3-0.6B and 1.7B no longer think by default, so replies come back faster (turn thinking back on with /think) • Removed the unstable VibeThinker-3B model (7B / 9B models cover the same ground)