Hugging Face

Hugging Face benchmarks not whether an agent solves a task with transformers but how much work it takes, sweeping models × library revisions × tasks on identical hardware to test whether a CLI, a Skill and self-contained examples actually cut agent effort.

Running a voice agent with nothing leaving the machine: llama.cpp serving Gemma 4 with explained flags, Silero VAD, Parakeet-TDT 0.6B for STT and Qwen3-TTS, assembled into a cascade that speaks a Realtime-API-compatible WebSocket and points a Reachy Mini at localhost.

IBM released granite-speech-5.0-2b plus revised 470m-turboctc and 4.1-2b-plus models. Changes include noise and reverb robustness, language-ID prompting, 2,000-term keyword biasing, better long-form translation and speaker attribution.

Three probes for benchmark contamination in speech recognition — consensus disagreement, masked entity retrieval, orthographic switching — run against 11 open models. Six reproduced VoxPopuli's erroneous reference transcripts over what the audio actually said; some keyed on acoustic cues identifying

NVIDIA published the base BF16 weights of Nemotron 3.5 Super VL 120B A12B, a multimodal hybrid MoE with MTP support. It is the pretrained foundation for the family's tuned and quantized variants.

NVIDIA details how Nemotron 3 models were specialized with SFT, RL and a generate-verify-refine loop to reach gold-level scores at IOI 2026 and IMO 2026, including per-stage gains and the role of test-time compute.

Thinking Machines' Inkling is out in the open: a ~975B-total, 41B-active multimodal MoE with 1M context that ingests text, image and audio, using relative attention rather than RoPE, shipping BF16 and NVFP4 weights plus MTP drafters, with a 276B/12B Small variant and day-0 vLLM, SGLang and llama.cpp

A Hugging Face author details how they prompted the ML Intern agent to train and publish six custom models. Covers prompt structure, baselines, smoke tests, spend caps, and a 0.8B distilled prompt rewriter built for $16.

InclusionAI's Ming-Image-0.1-Design-Layer decomposes a flattened design into a specified number of RGBA layers, evaluated on the Crello test set, complementing its Ming-Image-0.1-Design generation model.

InclusionAI ships Ling-3.0-flash-VL: a 124B-param, 5.5B-active MoE with up to 1M-token context, native image/video understanding, and agentic UI-interaction ability, scoring 42 on Artificial Analysis Intelligence Index v4.1.1.

Sentence Transformers v6.0 adds MultiVectorEncoder for ColBERT-style late interaction: one vector per token, scored with MaxSim, loading PyLate, Stanford ColBERT and colpali-engine checkpoints through the same API as dense models. Covers indexing, token pooling and page-image retrieval.

Amazon released ALoDLM-1.7B, a diffusion language model initialized from Qwen3 that adaptively loops shared transformer layers, keeps latent states for unresolved tokens, and uses a learned halting policy. Inference thresholds trade accuracy against throughput. Released under a non-commercial licens

Hugging Face's guide to Meta's Muse Spark Open: 256 experts with 8 active, 1M-token context, multimodal input, support in major inference engines, TRL fine-tuning and demos of browser agents.

One hf jobs run command brings up an OpenAI-compatible vLLM server on rented GPU hardware, reachable through an exposed port with your HF token as the bearer key — with notes on scaling to bigger models, SSH access, and driving a coding agent with it.

Ant Group releases Ling-3.0-flash-Fin, a 124B-parameter (5.1B active) MoE model continued-trained on financial data, with a 256K context window and benchmark results across FinFIRST, FinCRAFT, SpreadsheetBench, and tau3-Banking for financial agent workflows.

Hugging Face and AWS announce TorchNeuron, a native PyTorch backend for Trainium supporting eager mode, torch.compile and FSDP/TP, integrated with Transformers, Diffusers and Kernels, entering public beta on Oct. 20.

NVIDIA released Nemotron 3.5 Super VL 120B A12B in NVFP4 quantization. The hybrid LatentMoE vision-language model supports six languages and ships with Transformers, vLLM, SGLang and Docker serving instructions.

Ai2 details how it built AstaBrief 8B: 90K filtered real queries, SFT on citation-filtered reports, DPO preference pairs, and a single-pass report design, with open weights, data and a local example workflow.

Microsoft's ThinkingBox benchmark grades agents on the records they leave behind across 507 stateful MCP workflows, repeated 20 times to measure consistency and its cost. It runs through OpenEnv.

Search for research has to nail exact titles and fuzzy intent at once. The Papers with Code team walks through their hybrid stack — Postgres full-text plus pgvector joined by reciprocal rank fusion — and the storage and compute boundaries that keep embedding jobs off the request path.

NVIDIA Nemotron Labs released Dory1.0 4B, a small open-weights text-generation model trained on Nemotron pre- and post-training data. The BF16 checkpoint loads with Transformers and serves through vLLM and SGLang under an open model license.

Tencent released EVIE-8B, a 4096-dimensional late-interaction visual document retrieval model that tops the ViDoRe V3 leaderboard and serves as the distillation teacher for the smaller, compressible EVIE-4.5B.

Holo3.1 adds native function-calling alongside structured JSON so the models drop into third-party harnesses, pushes AndroidWorld from 67% to 79.3% on the 35B-A3B variant, and ships 0.8B/4B/9B sizes plus FP8, Q4 GGUF and NVFP4 checkpoints for local inference.

Hugging Face adds an RL Environments filter to the Hub. Environments are dataset repos of tasks, tests and reward rules, runnable through Harbor, Verifiers, OpenEnv and NeMo Gym, aiming to end siloed per-framework environment registries.