Models & Labs
IBM releases LLMON markup to fix prompt injection
Hugging Face Bloghigh confidence
Why it matters
- →Explicit structural markup reduces indirect prompt injection risks in small language models.
- →Special-token delimiters solve ambiguity issues inherent in JSON parsing for LLMs.
- →New Granite 4.1 preview models offer a practical path to more robust local inference.
IBM has released two experimental preview models based on Granite 4.1 (3B and 8B) trained to understand LLMON, an XML-like markup language designed for causal LLMs. The system uses reserved special tokens to explicitly tag instructions, data, and requirements, allowing the model to distinguish between user commands and retrieved content. Benchmarks show these models resist indirect prompt injection attacks significantly better than standard instruction-tuned models, with attack success rates dropping to near zero on specific tests. The release includes a GitHub toolkit for converting existing JSON pipelines to LLMON format.
Read originalThe story around this
Earlier coverage that leads up to this article, and what followed. Lines connect each piece to the closest one after it, converging here.
How we got here
- 1LLMs Vulnerable to Chain-of-Thought Forgery Attacks
MIT Technology Review AI · July 30, 2026 · Background
- 2Granite 4.2 LLMs Introduced by IBM and Hugging Face
Hugging Face Blog · August 25, 2026 · Related
- 3llama.cpp adds IBM ZDNN backend support
llama.cpp Releases · September 30, 2026 · Background
- 4llama.cpp fixes JSON schema enforcement in Ling 3.0
llama.cpp Releases · October 4, 2026 · Background
- 5Granite Switch 4.1.1 patches adapters without retraining base
Hugging Face Blog · October 29, 2026 · Background
More from Hugging Face Blog
© Hugging Face BlogModels & Labsother
Hugging Face releases TRELLIS.2 fine-tuning guide
Hugging Face finally provides a concrete path to adapt the powerful TRELLIS.2 model for specific domains like footwear or furniture, solving the 'generalist' problem in 3D generation. By targeting only the structure and shape DiTs with LoRA adapters while freezing the texture flow, they prevent catastrophic forgetting of material priors—a common pitfall when fine-tuning large generative models. This approach allows practitioners to achieve high geometric accuracy on small datasets (188 samples) without retraining the entire pipeline. It transforms TRELLIS.2 from a demo toy into a usable tool for production-grade 3D asset creation.
Hugging Face Blog
© Hugging Face BlogOpen Sourceagents
Tether AI launches BotPsy, a sub-billion agentic model
Tether AI Research has released BotPsy, an 0.8B-parameter model built on Qwen3.5 that specializes in tool-use and agentic workflows for on-device deployment. It achieves state-of-the-art performance in its size class, outperforming larger models like Qwen3.5-2B-Think on benchmarks such as BFCL v4 and ACEBench while maintaining strict privacy by keeping data local. The model is designed to handle routine tasks like calendar management and web search without relying on expensive cloud APIs, offering a viable path for private, self-hosted personal assistants. With open weights under Apache 2.0 and GGUF support, it provides immediate utility for developers building lightweight, autonomous agents.
Hugging Face Blog
© Hugging Face BlogResearchresearch
Pooling-Aware Training for Multi-Vector Embeddings
Multi-vector models like ColBERT are powerful but storage-heavy, often requiring hundreds of vectors per document. This technique teaches the encoder to preserve retrieval quality even when token vectors are aggressively merged via hierarchical clustering. By training on both full and pooled representations simultaneously, the model learns to compress information without losing critical distinctions. The result is a 4x reduction in vector count with only a marginal drop in accuracy, making large-scale semantic search significantly more efficient.
Hugging Face Blog
More in Models & Labs
Models & Labsother
llama.cpp b12340 adds CUDA 13 and ROCm 10.0
This release quietly closes the hardware gap for local inference by adding default support for CUDA 13 and ROCm 10.0 across Linux and Windows. Users with newer NVIDIA GPUs or AMD accelerators no longer need to compile from source to get performance, while Apple Silicon builds now ship KleidiAI disabled by default, shifting the focus back to standard ARM optimizations. The inclusion of Snapdragon support for Linux and Android marks a significant step toward on-device AI beyond the x86/ARM desktop ecosystem. It is less about new features and more about ensuring llama.cpp remains the universal runtime for every available accelerator.
llama.cpp Releases
Models & Labscoding
llama.cpp b12342 adds few-row MMQ for speculative decoding
Speculative decoding just got a significant speed boost on NVIDIA GPUs. This release introduces a specialized kernel for 'few-row' matrix multiplications, targeting the small batch sizes typical of verification passes in speculative decoding workflows. By optimizing these specific operations for quantized formats like Q8_0 and IQ4 variants, llama.cpp reduces latency where it matters most during token generation. The change is CUDA-only for now but leverages hardware capabilities from Ampere onwards, making local inference faster without requiring new model architectures or complex configuration changes.
llama.cpp Releases
© TechCrunch AIModels & Labsmodels
Google deploys Gemini 4 Hydrogen for faster Search
Google is quietly shifting its search infrastructure by replacing older fast models with Gemini 4 Hydrogen, a tier designed specifically for speed and cost-efficiency rather than raw reasoning power. This model completes tasks 2.5 times faster than its predecessor while beating the more powerful Argon on specific intelligence benchmarks, proving that specialized efficiency can outperform brute force in high-volume workflows. By integrating this into Search's fan-out technique, Google now breaks down complex queries into subqueries with greater precision and speed, delivering richer visual responses without the latency of heavier models. This marks a strategic pivot toward agentic execution layers within consumer search, making AI Mode feel snappier for everyday multi-part questions.
TechCrunch AI