What changed
0 fixes2 additions3 changes0 removals
- Performance
- Gameplay
- Compatibility
- Server
addedThis is a landmark update for Jingjie: for the first time, we have successfully embedded a local model directly into the system itself . With the first round of Path C stability fixes complete, and with the official integration of Ollama v0.20.5 and the Gemma 4 E4B local inference pipeline, “physical intuition” and “logical reasoning” can now form a stable loop directly on the player’s machine.
changedCore Upgrade: Local Model Pipeline Now Live Jingjie now includes Gemma 4 E4B as one of its core reasoning models. Combined with Flash Attention support in Ollama v0.20.5 , the model can deliver more efficient local responsiveness on compatible hardware. Its 128K context window also makes it better suited for complex research tasks, long-form dialogue, and structural analysis.
changedHardware Recommendations & Compatibility High-Performance Mode (Recommended GPU): On compatible systems with supported drivers and adequate GPU performance, Jingjie will enable Flash Attention for the fullest local inference experience. Base Compatibility Mode (CPU): Players without a dedicated GPU can still access local mode and use core features, though initial loading time and response speed will depend much more heavily on total system performance. Resource Note: Because this update includes model weights and the local inference environment, the download size is significantly larger. When entering related features for the first time, the system must initialize the Ollama service and load the model, which may result in a 10–20 second warm-up delay . After initialization, subsequent calls should become more stable.
changedStability & Fallback Fallback Protocol: If local model initialization fails, Jingjie will automatically fall back to the heuristic core or remote pipeline, ensuring that research flow is not interrupted. Path C Improvements: This update fixes UI jitter during multi-turn dialogue, restructures state synchronization for S2 / S3 , and improves focus stability and text selection reliability during interaction.
addedPlayers with GPUs will gain a deeper and more stable local-model experience, while players on lower-end systems now have a local mode that is at least runnable and resilient, rather than fragile or disconnected.
Jingjie: Signals changes
addedThis is a landmark update for Jingjie: for the first time, we have successfully embedded a local model directly into the system itself . With the first round of Path C stability fixes complete, and with the official integration of Ollama v0.20.5 and the Gemma 4 E4B local inference pipeline, “physical intuition” and “logical reasoning” can now form a stable loop directly on the player’s machine.
changedCore Upgrade: Local Model Pipeline Now Live Jingjie now includes Gemma 4 E4B as one of its core reasoning models. Combined with Flash Attention support in Ollama v0.20.5 , the model can deliver more efficient local responsiveness on compatible hardware. Its 128K context window also makes it better suited for complex research tasks, long-form dialogue, and structural analysis.
changedHardware Recommendations & Compatibility High-Performance Mode (Recommended GPU): On compatible systems with supported drivers and adequate GPU performance, Jingjie will enable Flash Attention for the fullest local inference experience. Base Compatibility Mode (CPU): Players without a dedicated GPU can still access local mode and use core features, though initial loading time and response speed will depend much more heavily on total system performance. Resource Note: Because this update includes model weights and the local inference environment, the download size is significantly larger. When entering related features for the first time, the system must initialize the Ollama service and load the model, which may result in a 10–20 second warm-up delay . After initialization, subsequent calls should become more stable.
changedStability & Fallback Fallback Protocol: If local model initialization fails, Jingjie will automatically fall back to the heuristic core or remote pipeline, ensuring that research flow is not interrupted. Path C Improvements: This update fixes UI jitter during multi-turn dialogue, restructures state synchronization for S2 / S3 , and improves focus stability and text selection reliability during interaction.
addedPlayers with GPUs will gain a deeper and more stable local-model experience, while players on lower-end systems now have a local mode that is at least runnable and resilient, rather than fragile or disconnected.
Operators,
This is a landmark update for Jingjie: for the first time, we have successfully embedded a local model directly into the system itself. With the first round of Path C stability fixes complete, and with the official integration of Ollama v0.20.5 and the Gemma 4 E4B local inference pipeline, “physical intuition” and “logical reasoning” can now form a stable loop directly on the player’s machine.
Core Upgrade: Local Model Pipeline Now Live Jingjie now includes Gemma 4 E4B as one of its core reasoning models. Combined with Flash Attention support in Ollama v0.20.5, the model can deliver more efficient local responsiveness on compatible hardware. Its 128K context window also makes it better suited for complex research tasks, long-form dialogue, and structural analysis.
Hardware Recommendations & Compatibility High-Performance Mode (Recommended GPU): On compatible systems with supported drivers and adequate GPU performance, Jingjie will enable Flash Attention for the fullest local inference experience.
Base Compatibility Mode (CPU)
Players without a dedicated GPU can still access local mode and use core features, though initial loading time and response speed will depend much more heavily on total system performance.
Resource Note
Because this update includes model weights and the local inference environment, the download size is significantly larger. When entering related features for the first time, the system must initialize the Ollama service and load the model, which may result in a 10–20 second warm-up delay . After initialization, subsequent calls should become more stable.
Stability & Fallback Fallback Protocol
If local model initialization fails, Jingjie will automatically fall back to the heuristic core or remote pipeline, ensuring that research flow is not interrupted.
Path C Improvements
This update fixes UI jitter during multi-turn dialogue, restructures state synchronization for S2 / S3 , and improves focus stability and text selection reliability during interaction.
ummary
Players with GPUs will gain a deeper and more stable local-model experience, while players on lower-end systems now have a local mode that is at least runnable and resilient, rather than fragile or disconnected.
The model is gradually evolving from an external tool into an internal organ of the Jingjie simulation.
JK Studio
2026.04.12
