webAI released TwIL-8B, a reasoning model built on IBM Granite-4.2-8B and optimized for math, code and instruction-following workloads, open-sourced under a non-commercial license. The lab claims best-in-group results on MATH-500, AIME 2025 and LiveCodeBench, reaching 88-93% of frontier-model capability on its strengths, while acknowledging limited world knowledge and no general-purpose scope.
Instruction Tuning
指令微调
Training method using instruction-response pairs to optimize model adherence to human intent
Curated Intelligence
webAI has open-sourced TwIL-8B, an 8.8B-parameter model built on IBM's Granite 4.2 8B and tuned for math, coding and instruction following. It reportedly fits in a 4.98 GiB quantized file that runs on CPUs or modest GPUs, scoring 81.6% on LiveCodeBench and reaching roughly 92-93% of frontier-model performance on coding and instruction following.
Baidu released ERNIE 5.5 Flash Base, a sparse Mixture-of-Experts language model with 280B total parameters and 14B activated per token, pre-trained from scratch on 33.5T tokens with a 128K training context and 1M-token support for fine-tuning and inference. The release is the end-of-pre-training checkpoint before any mid- or post-training, positioned as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research.
A new systematic study reveals that efficient activation steering methods for LLM conditioning often compromise generation fluency. It highlights a critical interaction where instruction tuning significantly reduces the effectiveness of steering compared to base models, while prompting and SFT are better at concept injection than removal. Additionally, cheap textual metrics were found to highly correlate with costly LLM-as-judge scores.
Alibaba's Qwen team released two new models: Qwen3-TTS-VD-Flash for voice design and Qwen3-TTS-VC-Flash for voice cloning. The former allows fine-grained control over timbre, emotion, and persona via natural language instructions, while the latter supports 3-second multi-language voice cloning. Benchmarks indicate superior performance over GPT-4o-mini-tts and ElevenLabs in InstructTTS-Eval and multilingual WER metrics.
Google has released the PaliGemma 2 Mix series on Hugging Face, featuring instruction-tuned vision-language models built on SigLIP and Gemma 2 architectures. Optimized for tasks like OCR and captioning, these models come in 3B, 10B, and 28B sizes with multiple resolution options, accompanied by a demo app and fine-tuning notebooks.
Hugging Face introduces SmolVLM-256M and SmolVLM-500M, with the 256M variant claimed to be the world's smallest Vision Language Model. These models maintain strong multimodal performance while significantly reducing footprint, offering direct support for Transformers, MLX, and ONNX across base and instruction-tuned checkpoints.
Addressing LLM vulnerabilities to prompt injection and jailbreaks, this research introduces the 'Instruction Hierarchy' concept. By training models to prioritize privileged instructions (e.g., system prompts) over user inputs, it effectively defends against malicious attempts to overwrite original directives.
InstructPix2Pix introduces a method for instruction-tuning Stable Diffusion to enable precise image editing via natural language commands. This approach allows users to modify images (e.g., "make the sky darker") without complex prompt engineering or per-image optimization. The core innovation lies in training on large-scale synthetic data to map editing instructions to visual changes.
This article demonstrates how to run the InstructPix2Pix (IF) model using the diffusers library on a free-tier Google Colab instance. It focuses on environment setup and basic inference workflows, serving as an introductory guide.