yuxinlu1 Gemma4-12B - Open-Source Coding and Agentic Model Series

What is yuxinlu1 Gemma4-12B

yuxinlu1 Gemma4-12B is an open-source coding and Agentic model series fine-tuned by individual developerLu Yuxinbased on Google Gemma 4 12B instruction models, comprising the V1 code version and the V2 Agentic version. Through distilling high-quality chain-of-thought data from Cursor Composer 2.5 and Fable 5 , the model compresses the coding reasoning capabilities of frontier large models into a locally runnable model with only 12B parameters, requiring a minimum of just 4.5GB of VRAM for offline operation, with no API fees, focusing on a private, local, zero-cost coding assistant experience.

Main Features of yuxinlu1 Gemma4-12B

• Three-tier model coverage for all scenarios: Sol targets complex Agent tasks and frontier research, Terra delivers GPT-5.5-level performance at 2x cost advantage, and Luna covers high-frequency lightweight work at the lowest cost.
• Dual enhanced reasoning modes: Sol adds a new "max" reasoning intensity, investing more time in deep thinking; "ultra" mode accelerates in parallel by invoking sub-Agents, breaking through the capability ceiling of a single Agent.
• Strongest safety stack and layered protection: Enhanced protection against high-risk activities, sensitive network requests, and repeated abuse scenarios, including model-level refusals, real-time classifier review, account-level risk signal monitoring, and differentiated access controls.
• Automated red teaming: Invested over 700,000 A100-equivalent GPU hours in automated red teaming, specifically seeking jailbreak attacks that generalize across multiple prompts and contexts, far exceeding the coverage of manual testing.
• Flexible prompt caching: Supports explicit cache breakpoints, with a minimum cache validity of 30 minutes; cache writes are billed at 1.25x the price of uncached input, while reads maintain a 10% discount.
• Cerebras hardware acceleration: Sol will be deployed on Cerebras chips in July, with inference speeds of up to 750 tokenper second, initially available to select customers.

Technical Principles of yuxinlu1 Gemma4-12B

• Tiered reasoning intensity architecture: Sol introduces OpenAI's highest reasoning intensity level to date, supporting the model in dynamically allocating more computation time for deep thinking on complex tasks, rather than a single fixed reasoning path.
• Multi-Agent collaborative orchestration (Ultra Mode): Through an upper-level scheduler, complex tasks are decomposed into multiple subtasks and distributed to different sub-Agents for parallel processing, breaking through the capability boundaries of a single Agent and accelerating overall task completion.
• Layered Safeguards：
  • Model layer: Trains the model to refuse prohibited cyberattack assistance, including intent disguise and jailbreak attempts.
  • Real-time layer: Deploys cyber and biological abuse classifiers to evaluate outputs in real time during generation; in high-risk scenarios, generation can be paused and the conversation context reviewed by a larger reasoning model.
  • Account layer: Flags activity to trigger cross-conversation account-level review, combining risk signals to distinguish persistent malicious behavior from legitimate dual-use security research.
  • Access layer: Differentiated access controls ensure that the most sensitive capabilities are not open to everyone by default, while preserving legitimate channels for defensive work.
• Automated red teaming and rapid response: Use own models to conduct large-scale automated red teaming, discover general jailbreak patterns; establish a rapid response process to reproduce, evaluate, prioritize, and fix newly discovered jailbreaks, and add them to the continuous evaluation library.
• Dynamic prompt caching mechanism: Supports developers in setting custom cache breakpoints within the conversation flow; the system charges a write fee at 1.25x the price of uncached input, applies a 90% discount when reading, and retains the cache for a minimum of 30 minutes.

Follow WeChat and reply "Open Source" to join theAI Open Source Project Exchange Group

How to use yuxinlu1 Gemma4-12B

• llama.cpp method: Download the GGUF quantized file that fits the device's VRAM from Hugging Face, install the latest version of llama.cpp that supports gemma4_unified and start the local service, then you can perform offline code generation and conversation through the browser WebUI.
• One-click client method: Import the downloaded GGUF model file in LM Studio, Jan, or Ollama, select the corresponding quantized version to load, enable Thinking mode, and directly enter programming or Agentic task instructions to interact.
• Transformers method: Pull the full-precision safetensors main weights from Hugging Face, load the model with the latest version of the Transformers library, and conduct secondary fine-tuning training or custom quantized deployment.

Core advantages of yuxinlu1 Gemma4-12B

• Extremely lightweight: 12B parameters + minimum 4.5GB VRAM requirement; it can run smoothly on consumer-grade graphics cards and Mac unified memory.
• Data quality first: The author emphasizes that data quality matters more than data quantity; about 10,000 execution-verified samples can achieve significant performance improvements.
• Zero-cost privacy: Runs completely locally, code is not uploaded to the cloud, and there are no API call fees, making it suitable for handling private projects.
• Significant performance leap: V2 improved from the base model's 15% to 55% on the tau2-bench telecom benchmark, an improvement of about 3.5x.

Project address of yuxinlu1 Gemma4-12B

• HuggingFace model library：https://huggingface.co/yuxinlu1

Comparison of yuxinlu1 Gemma4-12B with similar competitors

Application scenarios of yuxinlu1 Gemma4-12B