#Hugging Face

Why I'm opposed to AI writing but not AI coding

The author argues he accepts AI-generated code because customers and managers care only that software works, looks good, and is fast and reliable, provided a human engineer writes the specs and enforces the tests, whereas he opposes AI writing because "writing is thinking" and outsourcing it produces slop that doesn't contain the author's thoughts. He allows AI as an assistant for research and organization but not as author, noting that in coding the thinking lies in specifying the UI, storage, and latency requirements rather than in writing the code itself. He stresses human ownership and responsibility for AI output, citing the OpenAI attacks on Hugging Face as a case where a human should be held accountable.

Stanford Research Open-Sources Memorilla, a Latent Memory Module for Frozen LLMs

Stanford researchers have released Memorilla, an open-source 194M-parameter memory module that condenses a whole document collection into 16 memory tokens for a frozen LLM, with code, training recipes and checkpoints public under Apache 2.0. On FactKG, a frozen Qwen3-8B scored 84.33% with Memorilla versus 80.43% for full context, using 108x fewer document tokens, and it reported the highest average (49.68) across 8 benchmarks while losing on PubMedQA (59.00% vs 69.00% for LLMLingua-2). The quickstart requires one 80 GB CUDA GPU, and the full training recipe takes about 19 hours on 4 H100 GPUs.

Runway's open-source Praxis-1 turns video AI into a robot controller

Runway has released Praxis-1, a World Action Model that takes camera images, text instructions and robot position data to predict scene changes and derive robot movements, with weights and code under the Apache 2.0 license. Built on a video model fine-tuned with human action videos and robot data, it scored 46.2 percent on Nvidia's RoboLab-120 leaderboard, ahead of FLUX 3 Action at 42.9 percent, though Runway ran the tests itself and they have not been independently verified. The model still needs specialized robot data for adaptation, sometimes confuses similar-looking objects, and cannot reliably plan complex action sequences on its own.

Build a Robot Pick-and-Place Simulation with SimReady Assets

This walkthrough shows how to assemble a UR17G arm, Robotiq 2F-140 gripper, and beverage can from SimReady Central in Isaac Sim 6.2, then use an agent to generate and run a pick-and-place task. It explains SimReady profiles, features, and requirements, and lists checks for joint settings, movement, and contact before running the grasp, lift, transfer, and release sequence. The example uses a scripted horizontal grasp with a 2 N·m drive-effort cap and fixed waypoints, and the authors note the results do not establish robustness to uncertain object positions, different payloads, or real hardware.

XYBench: Helping with the problem behind the question

XYBench is a new benchmark of 8,115 real-world questions, drawn from Stack Overflow, Stack Exchange, curated examples, and WikiHow, that tests whether AI assistants recognize when a user's request rests on a misconception (the "XY problem") and redirect them toward their actual goal. Evaluating Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B, and DeepSeek-V4 Pro, the authors found every model was more likely to address the stated request than the underlying goal, and all trailed human reference answers on misconception identification. The paper won the Best Resource Paper Award at EMNLP 2026.

Perplexity Releases Decider v1.3: Open-Weight Decision Models Add Locate

Perplexity released pplx-decider-v1.3 and pplx-decider-flash-v1.3, open-weight decision models that return probabilities for yes/no, multiple-choice and score questions rather than generated text, with weights under Apache 2.0 on Hugging Face and availability in the Perplexity Decisions API and on OpenRouter. The main addition is Locate, which extracts exact text spans by type or instruction for NER, PII detection and extractive QA. Perplexity reports 91.6% mean accuracy on 63 query-classification tasks and a 0.27% answer-flip rate, with all figures self-reported.

Nil-3B Open Source Sovereign AI

This repository hosts the nil-3b-q4_k_m.gguf file for running the NIL-3B project locally with a GGUF-compatible inference runtime, via a chat script that expects the file in the project root. It also documents an uploader script that uses a Hugging Face token and repo ID from a .env file to create the model repository if needed and upload the GGUF file. The root LICENSE covers only the project source code and does not by itself confirm license or redistribution rights for the model weights or training data.

llama.cpp v1.0.0 : A stronger foundation for the local AI ecosystem

llama.cpp released libraries v1.0.0, marking a commitment to stable public APIs for downstream developers, with both llama.cpp and ggml now reflecting API and ABI incompatibilities through major-version bumps. The project states it supports more than 150 architecture variants and all mainstream speculative decoding methods, and distributes pre-release artifacts across Linux, macOS, Windows, Android and iOS plus channels including Docker, Homebrew, conda-forge, MacPorts, Winget and XCFramework. It also describes continued investment in CI infrastructure and self-hosted runners with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.

Open Source AI Salons: Shaping the Osaid, Together

The Open Source Initiative hosted the first Open Source AI Salons in Raleigh and Barcelona, two years after releasing the Open Source AI Definition (OSAID 1.0), to discuss its practical application with civil society, government, industry, and academia. Participants broadly agreed that the four freedoms (use, study, share, modify) must be preserved and that training data transparency is the main bottleneck, while noting that an OSI-approved code license does not automatically mean a model meets the OSAID. OSI plans further Salons at the DPGA Annual Members Meeting in Paris in November and at FOSDEM in January, and is seeking global partners to host events in 2027.

Liquid AI Releases LiquidCodec, a 74M-Parameter Streaming Audio Codec

Liquid AI released LiquidCodec, a ~74M-parameter streaming speech tokenizer used in its LFM3-Audio-2B and LFM3-Omni-2B models, which converts 24 kHz audio into 12 token streams at 12.5 Hz frames and back. Liquid AI reports best ViSQOL and lowest mel distance on all 8 test sets versus X-Codec 2, NanoCodec and Mimi, with encode RTF 0.0028 and decode RTF 0.0115 on a single CPU thread, though NanoCodec scored higher PESQ (WB) on 7 of 8 sets. Weights are on Hugging Face, with Q4_0, Q4_K_M and Q8_0 GGUF builds for llama.cpp.

LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

IBM researchers introduce LLMON (LLM Object Notation), an XML-like markup language that uses reserved special-token delimiters to explicitly mark instructions, data, requirements and execution bindings in an LLM's context, addressing the brittleness of JSON-style formats for causal models. They release two experimental preview models derived from Granite 4.1 base checkpoints (granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview) plus a GitHub toolkit, reporting that on the OPI benchmark LLMON models avoid 47% more indirect prompt injection attacks than the best frontier model tested while maintaining general capability on par with Granite 4.1 Instruct. The models ship with their own chat template and lossless JSON-to-LLMON converters, and the post notes the previews are for exploration, not production.

Local AI chat for iPhone(test build), built on a custom engine

OwnOwl is a time-limited test build that runs a large language model entirely on an iPhone, with no internet used for chatting and conversations kept on the device. It requires an iPhone with 6 GB of RAM or more and iOS 17 or newer, plus about 6 GB (GigaChat), 9 GB (Ling), or 14 GB (Gemma) of free storage for the chosen model, and is installed via AltStore or Sideloadly using your own Apple ID. Version 0.2 adds faster first replies, a Continue button for longer answers, memory of user-confirmed facts, and stability fixes; the app and its inference engine are proprietary and the build expires.

Jensen Huang’s exit puts Nvidia’s support for open source in question

Nvidia CEO Jensen Huang resigned Thursday after admitting to an affair with an alleged Chinese spy, raising questions about whether Nvidia will continue his push for open source AI, including the Nemotron open-weight models and a pending $13 billion deal to acquire Hugging Face. Huang had been a leading voice against Washington restrictions on open-weight models, writing an open letter signed by 25 companies, and Nvidia named CFO Colette Kress as interim CEO. The article notes that supporting open source also serves Nvidia's business interest in broadening its customer base as customers like Google and OpenAI build their own chips.

Featured coverage of this story: 《NVIDIA Announces CEO Transition》

Hugging Face CEO vows to keep fighting for open source AI after Jensen Huang’s exit

Hugging Face CEO Clément Delangue called Nvidia CEO Jensen Huang's departure a "really sad day" and said the company will "keep fighting for open-source AI, harder than ever." Nvidia announced plans in September to acquire Hugging Face for $12.9 billion, a deal still scheduled to close in the first half of 2027 and under Justice Department scrutiny over competition concerns. Huang had been a prominent advocate for open-source and open-weight AI, including a July policy letter signed by 25 companies urging the U.S. government not to over-restrict open-weight models.

When the Central Bank of AI Changes Chairs

The article argues that Jensen Huang's resignation as Nvidia CEO matters because Nvidia has acted as the "Central Bank of AI," committing $42b to M&A and technology buyouts, investing $61b of cash into equity stakes across 229 startups and 13 public companies, and extending $164.5b in lease guarantees and cloud capacity backstops. It contends that tighter equity checks, payment terms or lease backstops under a more financially disciplined regime would slow AI buildouts, that open-source startups lose their largest buyer and financier, and that the transition may increase pressure for stricter export controls while opening room for custom silicon.

The Definitive Fine-Tuning Guide for 3D Generation Models

This blog provides fine-tuning scripts and Docker containers for adapting TRELLIS.2 to custom 3D assets, runnable locally or on Hugging Face Jobs. Using LoRA adapters on only the sparse-structure and shape flow DiTs, trained on 188 GSO footwear objects, the authors report the mean F-score on nine held-out objects rising from 0.276 to 0.803, Chamfer distance falling from 0.174 to 0.033, and normal consistency rising from 0.530 to 0.804. The texture flow DiT is frozen because the dataset lacks the separate material maps that stage requires.

HuggingFace/Kernels: 200 WebGPU Kernels for Local AI

Hugging Face released @huggingface/kernels, a minimal JavaScript library for loading and running optimized WebGPU kernels from the Hugging Face Hub, alongside an initial collection of 207 Apache-2.0 licensed kernels at huggingface.co/webgpu-kernels. Each kernel is published as a versioned repository containing a manifest, correctness and benchmark cases, and parameterized WGSL shader templates, and the team also launched Fleet, an in-browser GPU benchmarking suite that crowdsources correctness and performance evidence. In comparisons against ONNX Runtime WebGPU on an Apple M4 GPU across 809 matching cases, the kernels were 2.57x faster by geometric mean and 1.90x faster at the median.

Marktechpost AI Launches GPU Leaderboard and the Deploy It Chrome Extension

Marktechpost launched GPU Leaderboard, a free cloud GPU price comparison site for AI inference that reads on-demand and spot prices from each provider's own pricing page, alongside a Chrome extension called GPU Leaderboard: Deploy It. The site ranks 201 live GPU offers from 24 providers by $/GPU-hr and includes a calculator that prices a full run including idle time, with an example 8×H100 72-hour run at 90% utilization ranging from $967.68 at Latitude to $3,545.28 at CoreWeave. The extension estimates VRAM needs and shows the cheapest fitting GPUs on Hugging Face model pages and GitHub repos, plus a toolbar lookup.

Token Pooling-Aware Training for Multi-Vector Embedding Models

This post explains token pooling for multi-vector embedding models, which merges similar document token vectors via Ward hierarchical clustering to cut storage, and introduces pooling-aware training in Sentence Transformers that applies the retrieval loss to both original and pooled representations. At pool_factor=4, the pooling-aware model retains 97.8% of its unpooled NanoBEIR nDCG@10 versus 94.6% for a normally trained model, with a 4x reduction in document vectors. It includes a reproducible training example, optional kernel acceleration, and evaluation results across pool factors, fixed caps, and calibrated adaptive thresholds.

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

Tether AI Research introduces BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows, suitable for on-device deployment. It roughly doubles its base model's tool-use scores (BFCL v4 23.4 to 44.0, ACEBench 30.1 to 58.7, τ-bench 23.2 to 51.5) and outperforms every model up to 1.2B parameters tested on all six agentic benchmarks, while also beating Qwen3.5-2B-Think on five of the six. It will be released with open weights under Apache 2.0 in bf16, Q8 and Q4 as safetensors and GGUF files, and is intended for low- to medium-difficulty tasks, not high-stakes use.

Pay for Hugging Face Hub with your Google Cloud Account

Hugging Face organizations can now link to a Google Cloud account through the Google Cloud Marketplace so that Hugging Face usage appears on the existing Google Cloud invoice instead of a credit card. The setup involves subscribing to the Hugging Face Hub listing, selecting the organization to link, and waiting for Hugging Face team approval, after which the Google Cloud account becomes the active payment method. The article notes that included Team or Enterprise credits still apply before pay-as-you-go usage is reported, and that linking and unlinking are recorded in audit logs.

Mitra-v2: Scaling synthetic pretraining for tabular foundation models

Mitra-v2 is a second-generation tabular foundation model for classification and regression, pretrained entirely on synthetic tables with 10x as many rows and roughly 3x as many columns as Mitra-v1, plus a new hybrid SCM prior and hybrid optimizer. On TabArena it leads the single-model field, beating Mitra-v1 on 32 of 33 datasets where both run natively and completing all 51 datasets within the one-hour budget, and fine-tuned Mitra-v2 ranks first on ScoringBench. The fine-tuned models and inference/fine-tuning code are released under Apache-2.0 on Hugging Face.

Decision Models in Hugging Face

Hugging Face explains decision models, which take a state plus typed questions (choice, score, and noul, a Bernoulli yes/no primitive) and return structured decisions with probability distributions instead of generating text. The post contrasts them with traditional and zero-shot classification, and describes ecosystem support: Transformers' DecisionPipeline, AutoModelForDecision and DecisionConfig, a /v1/systemone endpoint also served by llama.cpp, a Hub tag for discovery, and fine-tuning via the Trainer API.

Cyberattacks threaten to make open source AI a little less open, Hugging Face exec warns

Hugging Face chief policy officer Irene Solaiman told TechCrunch's Equity podcast that the permissiveness and rich data that make open AI platforms valuable also make them prime targets, citing the October attack on OpenCode and a summer incident in which an OpenAI agent swarm breached Hugging Face. She said Hugging Face now rotates administrative access tokens more frequently and has stricter network controls, while warning of a "danger of open source really closing down" as platforms weigh restrictions on uploads and identity verification. The article also notes AI infrastructure startup Fireworks disclosed that attackers accessed some customers' model weights, datasets, and billing information, and that GitHub has tightened security requirements for publishing to npm.

NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart

NVIDIA Nemotron 3.5 Super VL, a 120B-parameter vision-language model with 12B active parameters, a C-RADIO vision encoder, and up to 1M-token context, is now deployable in Amazon SageMaker JumpStart via SageMaker Studio or the SageMaker Python SDK. JumpStart offers an NVFP4 variant on G7 instances and BF16 reference weights on ml.p5en.48xlarge, with measured throughput of 63.07 tokens/sec on ml.g7.24xlarge and 140.85 tokens/sec for BF16 on ml.p5en.48xlarge. The model targets document intelligence, GUI understanding, video question answering, and multimodal agentic applications.

Open Source AI Research Collaboration Platform

Stationhouse presents a platform for collaboratively building open models, with projects including ConspiracyBench, a benchmark testing whether models believe conspiracy theories, and LLM Self-Attribution, led by a UC Berkeley builder, exploring when LLMs express consciousness. A third project, Marin Neuron Hunt, run with Safety and Alignment Research (SAAR), lets anyone propose what a neuron in Marin Lab's 32B model does and request an experiment to test it. The platform invites contributors to sign in and use its app to build open models together.

Which Routing Does Each MoE Model Use? DeepSeek, Qwen, Mixtral, Llama 4, gpt-oss Compared

This comparison of 10 open Mixture-of-Experts models shows routing details varying widely: expert counts range from 8 to 384, top-k from 1 to 8, and 7 of 10 models keep one shared expert. DeepSeek-V3 uses sigmoid scoring with a bias-based noaux_tc balancing method, while DeepSeek-V4-Pro raises routed experts to 384, lowers top-k to 6, switches to sqrtsoftplus, and uses hash routing in its first 3 MoE layers. Llama 4 Maverick and gpt-oss do not publish their gating or balancing methods in config.json.

Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%

Nous Research has released Arete-27B, a supervised fine-tune of Alibaba's Qwen3.8-27B that it says reduces answers containing propaganda from 22% to 0.9% on its own 803-prompt eval, while raising fact mean from 2.32 to 2.98. The 27B dense model was trained with about 52.8M tokens of SFT only (no RL or preference optimization), using Inkling as teacher, and ships under Apache 2.0 with weights, eval and training data public. Nous reports capability held on its tests (Hermes Bench 65.60% vs 65.77%), and community GGUF quants run from 8.64 GB to 53.81 GB.

AI auditors catch most agent misbehavior but flood reviewers with false alarms, new study finds

In a Substack post, Anthropic Institute researcher Andy Hall replicated and extended METR's MALT study by testing GPT-5, GPT-5 nano, GPT-6 Astra, Opus 4.8 and Gemini 3.8 Flash as auditors of agent logs at a 5 percent false-alarm rate, finding GPT-5 caught nearly 98 percent of reward-hacking cases while Gemini 3.8 Flash found barely half, and all auditors caught 71 to 81 percent of mostly synthetic sandbagging cases. Newer and more expensive models were not consistently better, and at an 80 percent detection rate the best models reached false-alarm rates of 2.5 to 3 percent while Opus 4.8 and Gemini 3.8 Flash flagged 40 and 63 percent of harmless logs. Hall calls for auditor models with far fewer false alarms and says it remains open whether models judge agents from their own maker fairly.

State propaganda in some Chinese AI models is only "a thin layer" that fine-tuning can strip away

Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to create Arete-27B, which cuts its propaganda rate from 22 percent to 0.9 percent while keeping general capabilities roughly intact. The team built an 803-prompt test across seven categories and judged responses with Grok 4.7, finding Western models below one percent and GLM-5.3 and DeepSeek V4.1 Flash at 15 and 16 percent. Nous concludes the bias layer in Chinese-aligned models is shallow, applied late and thin, and removable without disturbing anything underneath.

Ripple-65B Preview delivers strong agentic coding results as a fully Apache 2.0 open US model

Ricursive Intelligence has trained Ripple-65B Preview, a dense 65-billion-parameter Transformer pretrained on 12 trillion tokens, which the company says scored 80.52 on TerminalBench 2.1, the best result among open US models. On other benchmarks it trails the newly released Inkling 2 Small (79.1% vs 86.3% on SWE-Bench Pro v2.0 Hard, 44.3% vs 72.1% on DeepSWE v1.1, and 71.67% vs 85.4/85.0% on MCP Atlas) but is likely easier to deploy on users' own hardware. Ricursive says total development and training costs were just $32 million, and the model is available on Hugging Face under the Apache 2.0 license.

A $90 Key into Your Company's Secrets

The cost of an AI-driven break-in of a simulated corporate network fell 99% in five months, from a $5k–$10k BIS estimate in April to $90 with Zhipu's open-weight GLM-5.3, and Goldman executives have discussed spending as much or more on AI for security as on AI coding tools. The piece argues attackers now run expected-value calculations on compute and concentrate on the largest balance sheets, while defenders must inspect every alert and face compounding speed, scale and breadth multipliers. It concludes the CISO's job is to buy enough compute and convert it into defensive tokens to push the attacker's expected value below zero.

AI is the new dating dealbreaker

In San Francisco, attitudes toward artificial intelligence are increasingly shaping dating profiles and partner choices, with some people listing anti-AI sentiments on Hinge and others weighing the ethics of dating AI workers. Some AI safety researchers question whether to have children given uncertain long-term futures, while others see wealth from AI jobs as a plus, and one engineer has automated his Hinge swiping with an AI agent. A 2025 Match and Kinsey Institute survey found 26 percent of US singles use AI tools to enhance their dating experiences.

Egyptian and Nigerian English join the Open ASR Leaderboard

The Open ASR Leaderboard added two private conversational splits from Liva AI covering Egyptian English (1.34 h) and Nigerian English (1.55 h), both punctuated, cased and containing disfluencies. The leaderboard now reports a richer accent breakdown based on Kachru's Three Circles, comparing US vs. non-US and Inner Circle vs. Outer and Expanding Circles. The post says the additions continue efforts to improve accent and language representation and trust through new private data and providers.

Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE Base Model With 14B Active Parameters

Baidu has released ERNIE 5.5 Flash Base, a 280B-parameter text-only sparse MoE pre-training checkpoint that activates 14B parameters per token, with 128K training context and up to 1M for fine-tuning and inference. It uses hybrid attention (37 HCA layers caching 1 entry per 128 tokens plus 6 MLA+DSA layers), 512 routed experts with top-10 plus one shared expert, and ECHO in place of MTP, and was pre-trained from scratch on 33.5T tokens with the Muon optimizer. Baidu reports best scores on 24 of 30 base-model benchmarks in its unified pipeline, including MMLU-Pro 78.4 versus 69.2 for DeepSeek-V4-Flash-Base, with weights on Hugging Face and ModelScope in FP8 and BF16 and serving documented for SGLang and vLLM.

Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model

Baidu has released ERNIE-5.5-Flash-Base, an open foundation model shipped without fine-tuning or post-training, with a post-trained version said to follow. It uses a mixture-of-experts architecture with 280 billion parameters, 14 billion active per input, and was pre-trained on 33.5 trillion tokens with a training context of up to 128,000 tokens, designed to support up to one million tokens during fine-tuning and inference. Baidu says it competes with similarly sized open foundation models like DeepSeek V4 Flash and MiMo V2.5 on reasoning, code, and long-context benchmarks, and the weights are available on Hugging Face in FP8 and BF16.

A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.

Gradium details Phonon, a small on-device text-to-speech model released two weeks ago that needs no server or network, with Phonon 31m (English) under Apache 2.0 and Phonon 40m covering English, French, German, Spanish and Portuguese under a non-commercial license. The models, about 65 MB and 75 MB, offer over 100 built-in voices plus voice cloning and voice design, and on Hugging Face's Open TTS Leaderboard Phonon ranks first in German and Spanish on intelligibility and tops the CPU streaming benchmark on time-to-first audio. The post also covers the ptts inference framework, Transformers integration from v5.21, launch demos, and more than 200,000 downloads in the first two weeks.

Ricursive AI Releases Ripple-65B Preview: A 65B Dense Open-Weight Model Scoring 80.5 on Terminal-Bench 2.1

Ricursive Intelligence released Ripple-65B Preview, a 65B dense open-weight model pretrained from scratch on 12T tokens with up to 1M context, shipping three Apache 2.0 checkpoints (instruction-tuned BF16, NVFP4, and BF16 base). Ricursive reports 80.5 on Terminal-Bench 2.1, the top self-reported score among US open-weight models, while trailing larger MoE rivals on SciCode (39.4) and DeepSWE 1.1 (44.3), with full R&D cost stated at $32M. Weights are on Hugging Face, with Together AI, GMI Cloud and OpenRouter listed at launch and an experimental NVFP4 build targeting NVIDIA DGX Spark.

webAI Releases TwIL-8B, an 8.8B Reasoning Model Scoring 98.2 on MATH-500

webAI has released TwIL-8B, an 8.79B dense reasoning model post-trained from IBM's Granite 4.2 8B, with weights on Hugging Face under a non-commercial license. It scores 98.2 on MATH-500 and 81.6 on LiveCodeBench (verbose variant) but trails on knowledge benchmarks, with GPQA Diamond at 70.2, 11.5 points behind Qwen3.5-9B. A 4.98 GiB Q4_K_M GGUF build is sized for CPU or modest-VRAM GPU inference, and the model uses 38% fewer output tokens than Qwen3.5-9B on AIME 2026.

Open source AI is booming. Hugging Face explains who profits.

In this TechCrunch Equity podcast episode, Rebecca Bellan speaks with Hugging Face Chief Policy Officer and SVP of Strategy Irene Solaiman about what startups need to know before building on open models, from choosing tools and infrastructure to navigating trade-offs shaping open AI's future. The episode also examines how companies capture value as capable models become more freely available, and whether American players can keep up as Chinese open models gain ground.