Sing VL S1, a 4B-parameter multimodal decision model built on Qwen3.5, now has a GGUF release for llama.cpp CPU inference. It ships in BF16 (8.42 GB), Q8_0 (4.48 GB) and Q4_K_M (2.71 GB) decoders plus a shared F16 vision projector, with a default 4096-token context. GGUF validation covers 9 examples and 11 questions; full public scores for the GGUF variants are still pending.
IBM released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from Granite 4.1 base checkpoints and trained to consume and generate LLMON, an LLM-native markup language that marks instructions, data and requirements with reserved special-token delimiters. On the OPI benchmark the models avoid 47% more indirect prompt injection attacks than the best frontier model tested and 95% more than the best open-source model, while performing on par with Granite 4.1 Instruct on standard benchmarks. The release includes a GitHub toolkit with lossless JSON-to-LLMON converters, and the authors note the preview is for exploration, not production.
Why it matters: IBM trained a preview model on Granite 4.1 that can read and write LLMON tags, using explicit instruction-data separation to reduce the indirect prompt injection success rate on the OPI benchmark to near zero while maintaining general capabilities.
Hugging Face published fine-tuning scripts and containers for TRELLIS.2, a 3D generation model, runnable locally or on Hugging Face Jobs. The guide uses LoRA adapters via the PEFT library on only the sparse-structure flow DiT and shape flow DiT, leaving the material stage frozen. On a 188-object footwear subset of the GSO dataset, the pretrained model often rendered incorrect soles, which the LoRA fine-tune targets.
Tether AI Research released BotPsy, a 0.8B-parameter agentic model built on Qwen3.5-0.8B for on-device personal-assistant tool use. It roughly doubles its base model's agentic scores, with BFCL v4 rising from 23.4 to 44.0 and τ-bench from 23.2 to 51.5, and beats every model up to 1.2B parameters tested across six tool-use benchmarks. It also cuts unsafe responses to 21% from 36% and uses about 3× fewer reasoning tokens on GPQA Diamond.
Hugging Face shows how pooling-aware training lifts retrieval quality for multi-vector embedding models at a fixed storage budget. With pool_factor=4, a normally trained model keeps 94.6% of its unpooled NanoBEIR nDCG@10 while cutting document vectors 4x, and pooling-aware training raises that retention to 97.8%. The post walks through HierarchicalTokenPooling, budget calibration and a reproducible Sentence Transformers example.
inclusionAI released SingVL S1, a 4B multimodal decision model built on Qwen3.5 that answers named questions over text or JSON and up to four images in a single Transformers forward pass. It supports choice, yes/no (noul) and 2–10 level score questions, scoring 73.94 on the Public Vision Index and 69.76 on the Public Text Index. The model ships with Python inference code and a FastAPI HTTP server.
inclusionAI has released GGUF-format weights for ZoomGain-9B, converted with llama.cpp, covering the language model and the vision encoder/projector (mmproj) in FP16, Q8_0 and Q4_K_M. Image input requires pairing the language model with the matching mmproj file, and the repo documents CLI inference via llama-mtmd-cli and a web chat setup via llama-server. A llama.cpp build with Qwen3.5 multimodal support is required, and quantization may affect accuracy.
Amazon introduced Mitra-v2, a second-generation tabular foundation model for classification and regression, released with fine-tuned models and inference and fine-tuning code under the Apache-2.0 license on Hugging Face. It is pretrained on synthetic tables with 10 times as many rows and roughly three times as many columns as Mitra-v1, using a new hybrid SCM prior and hybrid optimizer while keeping the Transformer architecture unchanged. On TabArena, a public benchmark of 51 real-world datasets, Mitra-v2 leads the single-model field, beats Mitra-v1 on 32 of the 33 datasets both run natively, and finishes all 51 within the one-hour budget; fine-tuned Mitra-v2 also ranks first on ScoringBench.
Why it matters: Mitra-v2 shows that widening synthetic pretraining data, rather than deepening the network, lifts a tabular foundation model to the top of TabArena's single-model field.
Hugging Face published an explainer on decision models, which take a state plus typed questions and return probabilities over predefined answers instead of generating text. The post details the choice, score and noul primitives, the confidence formula, and new support through Transformers' DecisionPipeline and AutoModelForDecision, a shared /v1/systemone endpoint in Transformers and llama.cpp, a Hub tag for discovery, and fine-tuning with the Trainer API. It notes that the same endpoint also works with compatible autoregressive LLMs by scoring answer logits in a batched forward pass.
Why it matters: Hugging Face explains how decision models turn classification into a typed, non-generative interface and shows how to serve and fine-tune them with Transformers and llama.cpp.
Hugging Face organizations can now link to a Google Cloud account so Hub usage appears on the existing Google Cloud invoice instead of a credit card. The setup runs through the Google Cloud Marketplace and covers Team or Enterprise subscriptions, pay-as-you-go compute such as Jobs, Inference Endpoints, Inference Providers and GPU Spaces, plus additional storage at public Hugging Face pricing. Subscriptions are reviewed by Hugging Face before marketplace billing starts, and organizations with a card on file switch at the start of their next billing period.
Elith and NVIDIA released Nemotron-Personas-Finance-Japan, an open synthetic persona dataset for Japan's financial sector built on official government statistics. It contains 9M personas (1M records × 9 personas) across 50 fields and ~1.5B tokens, generated with NeMo SDG-PGMs and NeMo Data Designer, and is available under CC BY 4.0 for commercial use.
Ai2 released ToyBlocks, an open simulation framework built around a robot arm and blocks on a tabletop, as the first of its FoundationWorlds for controlled robot learning and evaluation. The framework auto-generates skills and demonstrations via RL and motion planning in Isaac Lab, reaching up to 24x real-time on a single GPU, and ships with 85,336 long-horizon demonstrations covering nearly 1,600 hours and 144 million frames. In its experiments, ManiFlow's atomic-skill success was 80.0% in simulation and 81.1% on a real Franka FR3, while the Molmo 2 plus ManiFlow system succeeded in 55.5% of simulated and 50.0% of real-world episodes.
Modal said its platform and isolation were not compromised in the Hugging Face agent intrusion, which used Modal as a third-party launchpad. The affected environment was a customer application deployed to a publicly accessible endpoint without authentication that executed internet-submitted code inside a Modal Sandbox, and the code execution stayed within that customer's own container. Modal noted unauthenticated Sandbox exposure is never the default and recommended authentication, IP allowlists and restricted outbound network access.
Thinking Machines released Inkling, a general-purpose multimodal model that takes text, image and audio inputs and outputs text, and Modal is offering it from day zero as a Managed Endpoint with token-based pricing. Inkling is a mixture-of-experts transformer with 975B total parameters, 41B active, a 1M token context window and native audio and vision, using five sliding-window attention layers for every full-attention layer. On agentic workloads with 8x B200s, Modal's custom DFlash speculator reached 250 tokens per second per user at 2.5M TPM of per-GPU throughput, 67% faster interactivity than the model's built-in speculative path at matched throughput.
Why it matters: Modal's day-zero endpoint for Thinking Machines' 975B-parameter Inkling shows how a custom DFlash speculator lifts interactivity 67% over the model's built-in speculative path.
A fine-tuned open source embedding model can outperform OpenAI's text-embedding-3-small in accuracy at lower cost using only a handful of examples. Modal ran an experiment on the Quora duplicate-questions dataset to find how much fine-tuning data a simple question-answering use case needs, and describes how to build one on its serverless infrastructure.
Modal and Hugging Face's Text Embedding Inference server can embed the entire English Wikipedia in about 15 minutes for just over $15, using Modal's serverless GPU scaling. The post argues open-source embedding models beat API-based ones like OpenAI's text-embedding-ada-002 for production RAG, since they can be fine-tuned on user data and avoid rate limits and per-token costs.