fix: update model

fix: remove exAI

fix: still error

fix: still error

fix: chat error

fix: chat error

fix: check_model_input issue

fix: update app py

fix: update req

fix: update req

fix: update req

fix: update req and docker

fix: update base model

fix: exAI ropeparam

fix: update req

fix: for exai req update

fix: update req and docker

fix: update dockerfile

fix: update req and head method

fix: update dependency

fix: update requirements for model

fix: update backend lib with log

fix: update model lists

fix: model changing

feat: select model from admin panel

fix: update model to 360m

feat: update base free model

fix: update tier mode

fix: update limit

feat: add ping for uptime

fix: update context window and prompt

feat: add cerebras api and super mode

feat: answering question

feat: line number and multi docs

fix: improve chunking

fix: upload issue large than 23 mb

fix: llm model info

feat: add groq api and deepmind

perf: switch default LLM to SmolLM2-1.7B - 40-50% faster tok/s, better instruction following

fix: JSON-encode SSE tokens to preserve newlines in markdown; reduce top_k to 3

perf: greedy decoding + dtype fix - 2-3x faster inference on CPU

fix: rewrite loader.py as clean UTF-8 - remove Windows-1252 em-dashes causing SyntaxError

feat: self-hosted Qwen2.5-1.5B-Instruct via transformers — no external API, no compilation

feat: replace llama-cpp-python/Groq with free HF InferenceClient (zero compilation)

fix: restore build-essential+cmake, pin llama-cpp-python==0.3.16 for stable layer cache

fix: use pre-built llama-cpp-python CPU wheel — eliminates 8min C++ compile

fix: Dockerfile — pre-install CPU torch, upgrade llama-cpp-python to >=0.3.14 (qwen3 support)

fix: upgrade llama-cpp-python >=0.3.14 for qwen3 arch support (was 0.3.8, pre-May 2025)

fix: double .gguf extension — skip symlink when path already ends in .gguf; add verbose step logging

feat: LLM readiness tracking — 503 while loading, llm_ready in /health