fix: Switch ERNIE 21B IQ2_XXS to bartowski repo

fix: Correct IQ2_XXS filename pattern for ERNIE 21B

feat: Add ERNIE 21B IQ2_XS variant - more stable quantization

feat: Add thinking buffer for reasoning models - extends n_ctx by 50% of max_tokens when reasoning enabled

fix: Make download component visible for JSON export

fix: Download button now works - removed broken .then() chain

fix: Disable download button until content exists - prevent empty download field UX issue

feat: Add comprehensive generation metrics to JSON export

Fix: Move update_reasoning_visibility to module level - was accidentally defined inside load_model function

Invalidate Docker cache to force fresh app.py copy

Fix: Remove duplicate return statement in get_model_info

UI/UX improvements: language selector, dynamic reasoning, copy buttons, adjustable temperature

Add bilingual support: English and Traditional Chinese (zh-TW)

Update CLAUDE.md with current implementation details

Adjust ERNIE-4.5 21B Thinking generation settings for better summarization

Update AGENTS.md with Gradio app coverage and deployment info

Add ERNIE-4.5-21B-Thinking (Q1_0) to model registry

Reorder models by total parameter count (smallest to largest)

Add 3 new models: Granite 4.0 Tiny 7B, LFM2 2.6B Transcript, Qwen3 4B Thinking

Add model-specific inference settings and dynamic UI controls

Remove DeepSeek-R1 1.5B and add repeat penalty to prevent loops

Add 3 new 1.5B/1.7B models with Q4 quantization and reasoning toggle

Upgrade model precision to Q8_0 for better quality

Remove Qwen3 Q2 model (low quality)

Add multi-model selection with CPU optimizations

Fix: Restore tag-based thinking parser, remove slow marker scanning

Major UI/UX improvements - Enhanced interface with modern design

Major UI/UX improvements

Fix: Stream thinking tokens live instead of showing all at once

Fix: Separate thinking and summary fields during streaming

Fix: Support both <think> and <thinking> tag formats in parser

Test: Verify git workflow preserves commit messages

Test commit: verify git push workflow to HF Spaces

Add CLAUDE.md with comprehensive repository documentation

Fix error handler yield types to match Gradio outputs

Fix streaming: show all output in thinking field, parsed in summary

Stream thinking token-by-token in real-time

UI improvements: separate thinking/summary fields + markdown rendering

Add missing parse_thinking_blocks function

Remove show_copy_button parameter (not supported in Gradio 5.x)

Fix streaming output issues

Add model info display to UI

Increase context window from 4096 to 32768 tokens

Fix runtime error: add libgomp1 for OpenMP support

Use prebuilt llama-cpp-python wheel for faster builds

Exclude transcripts from HF Spaces deployment

Add HuggingFace Spaces demo with Gradio UI

comprehensive model benchmark: 6 models evaluated for transcript summarization

separate Qwen3 thinking blocks into thinking.txt and summary.txt

add to zh-TW post-processing