Meituan has released the weights for LongCat-2.5 under the MIT License. The native multimodal MoE is built for long-horizon coding and agent tasks, with 1.6T total parameters and about 48B activated per token. 🤖 • Stronger coding agents. LongCat-2.5 improves substantially over LongCat-2.0 on DeepSWE and Terminal-Bench v4.0 in Meituan’s tests. • Visual coding. The model processes image and text tokens together, so it can use a screenshot as a reference when writing webpage code. • 3D modeling through code. The model can turn a text description into a Blender script and leads the models in Meituan’s 3DCodeBench Text-to-3D comparison. • Training for longer tasks. Agentic RL uses real engineering tasks and reproducible agent failures. LongCat-2.5 also posts the highest BrowseComp score in Meituan’s reported comparison. • INT4 inference on NPUs. With quantized KV cache, W4A4 nearly doubles decode throughput over W8A8KV16 in Meituan’s stress test.

Adventists AI just open-sourced Sophon-Realtime v1 on ModelScope: a full-duplex voice model that hears how you speak, and answers with tone of its own. • It keeps listening while you talk, answers ~0.4 s after you stop, and stops the moment you cut in • It laughs, sighs and clears its throat when the conversation calls for it • #1# among open full-duplex models on VoiceBench (82.8) and Big Bench Audio (81.8%) • 48% Pass@1 on Full-Duplex-Bench v3 tool calling, ahead of Grok (43%) Apache-2.0 weights. Built by @BKJinAI and team.

THX-02 from @halx_ai is open source: a 4.5B multimodal decision model trained on requests up to 32K tokens, with a 60M adapter for selective thinking. It returns probabilities over your supplied options. 🤖 🧠 Reason selectively: Auto mode generates a short reasoning note when the direct answer is uncertain and an internal gate identifies a need for reasoning, then combines the direct and thought-assisted predictions. ⚡ Fast direct decisions: Median latency is 50 ms directly and 79 ms in auto mode on one RTX PRO 6000. Requests that trigger thinking typically take 0.3–1 second. 📊 Measured gains from thinking: Auto mode raises Decision Index 0.3 from 63.83 to 65.06 on a 6,018-request public-suite sample. Full-suite and private leaderboard evaluations remain pending. 🎯 Define your own decisions: Ask up to 64 questions per request using multiple-choice, yes/no, or ordered ratings. New labels need no retraining. 🖼️ Text, images, and charts: Accepts up to four images per request, with post-training and evaluation across more than 20 languages. 📜 Apache 2.0.

Baidu is opening ERNIE-5.5-Flash-Base straight out of pre-training: an open-weight foundation for domain adaptation, instruction tuning, and RL research. 🤖 • 280B total parameters; 14B activated per token • Pre-trained on 33.5T tokens, with training context up to 128K • Up to 1M-token context for fine-tuning and inference • FP8 and BF16 weights 🛠️ A starting point for your own training: The checkpoint precedes all mid- and post-training, leaving developers to choose their own adaptation methods and objectives. 📊 Strong base-model results: Scores 78.4 on MMLU-Pro, 82.3 on HumanEval+, and 40.1 on LiveCodeBench, ahead of the three compared base models on these benchmarks under a unified evaluation pipeline. 🧠 Compressed attention for long inputs: Combines 37 HCA layers, each caching one entry per 128 tokens, with six MLA-DSA layers for full-context access through sparse attention. ⚙️ A built-in speculative draft: ECHO reuses the backbone's final layer to predict future tokens, replacing a separate MTP module. 📦 Deployment: Transformers, SGLang, and vLLM-ERNIE paths are available. The Base model uses the completions endpoint.

EMLab's KnowLine offers seven open-weight decision models for RAG, agent harnesses, and enterprise workflows, spanning dense models from 0.8B to 31B plus two MoE options. 🤖 🏆 SOTA on private decision tests: KnowLine-31B-Gen1 achieves 71.86 on Jev Decision Index 0.4.1 Full. The score combines private same-skill and new-domain tests, excluding public-suite results. ⚡ Low latency with sparse activation: KnowLine-35B-A3B-Gen1.5 activates about 3B parameters and scores 65.77 Full, with 47.1 ms median latency on the leaderboard's single RTX PRO 6000 setup. ✅ Measured probability calibration: In the 31B model's public-suite evaluation, approximately 0.9–1.0% of decisions made with at least 95% confidence were wrong. Validate thresholds on your own data. 🎯 Decisions throughout the workflow: Filter retrieved documents, check supporting evidence, select tools, route tickets, and score outputs through a Jev-compatible API supporting up to 64 questions per request. 🛠️ Adapt to your business: Continue post-training with ms-swift. The development loop uses research agents to construct and decontaminate data, train, evaluate, and retain or roll back changes based on evidence. 📜 Weights: Apache 2.0. Serving code: MIT.

AgentScope introduces Piccolo: open-weight decision models in 27B, 9B, and 4B sizes for agent and data workflows. Piccolo evaluates candidate options in context and returns structured decisions with probabilities. 🤖 🌐 Explore Piccolo: 🎯 Decisions that drive action: Control game characters, select and review tools, retain relevant context, filter data, assess output quality, and choose the next step—all through one decision interface. 📊 Benchmark results: Piccolo-27B achieves 92.21% accuracy on JevBench Public and scores 68.59 and 77.75 on text and multimodal Decision Index Public, respectively. On our purpose-built benchmarks, it achieves 84.41% on HarnessJevBench for agent execution control and 82.60% on DataJevBench for data workflows. 🔄 Recursive self-improvement: Model errors and per-example learning signals guide data selection, synthesis, and changes to training and inference. Research agents test each change and use the results to plan the next round, within human-defined goals and budgets. 📜 Apache 2.0; third-party licenses apply. 💻 Code:

At the physical workspace of ModelScope, China's largest AI open-source community, a single developer can do the work of a team. Read more in #XinhuaHeadline#: World Internet Conference Wuzhen Summit highlights China's open-source AI push

Coding, 3D creation, and financial research with open weights. Kimi K3.1 Pro is Moonshot AI's multimodal MoE with 2.8T total parameters, 104B active, and a 1M-token context. 🤖 💻 Stronger coding execution: Scores 65.7 on Terminal-Bench 4.0. On the internal Kimi Code Bench 3.0 Long-horizon, it improves from K3's 55.95 to 61.2. 🔎 Research with tools: Reaches 70.0 on HLE-Full with tools and 92.1 on BrowseComp in the published evaluations. 🧊 Build and refine 3D assets: Writes and runs Blender and CAD scripts, scoring 92.7 Voxel IoU on BenchCAD with tools. 🏆 Financial document analysis: Achieves SOTA on OfficeQA Pro v2 at 64.4, covering retrieval, cross-document verification, and multi-step numerical reasoning. ✅ Fewer unsupported claims: Claim-level hallucinations fall from K3's 23.4% to 8.3% on the internal Kimi Hallu Bench Hard with tools, a stress test rather than a measure of everyday error rates. 📜 Kimi K3.1-Pro License; commercial-use and attribution conditions apply.

Baidu's MarathonDocBench evaluates the parsing of 100–300-page Chinese and English documents, including cross-page text, tables, and document structure. 🤖 📚 Hundred-page scale: 100 documents, approximately 18,000 pages, and 205,000 annotated layout elements. 🔗 Structure across pages: Human-refined annotations cover bounding boxes, heading hierarchy, reading order, and cross-page text and table continuations. 🧪 Document-level evaluation: Reuses MPDocBench-Parse's eight metrics to assess text, tables, formulas, reading order, figures, and heading structure. Overall is their mean. 📊 A demanding test: Listed general-purpose VLMs score below 34 Overall under the published whole-document protocol, which counts failed requests as empty predictions. Input and output limits also affect results. 📜 Apache-2.0 License.

Describe a scene, then build it in Three.js. Qwen3.5-4B-3D-Coder is CCORE's open-weight fine-tune of Qwen3.5-4B for interactive 3D scene generation, available as an early preview. 🤖 🧊 Generate scene code: Create editable HTML and JavaScript for architecture, interiors, gardens, and voxel environments. 🖼️ Use visual references: Q4_K_M and Q8_0 GGUF versions support image-guided generation with the original Qwen3.5-4B vision projector. Outputs interpret the reference rather than reproduce every detail. 💻 Run locally: FP16 weights support Transformers; GGUF versions include llama.cpp setup instructions. 📚 Explore the training data: 4,146 scene examples include prompts, reasoning traces, and complete HTML/JavaScript responses. 📜 Model: Apache 2.0. Dataset: MIT.

With uni-robot-agent, open-weight Qwen shows robot-use capabilities comparable to the published Astra–RoboProbe results on RoboDojo. The unified harness lays the groundwork for generalizable agentic training. 📄 🏆 Stronger robot use: Qwen3.8-Flash-Next achieves 31.39 Score / 25.00% success without privileged simulator inputs, versus 28.97 / 22.48% for the published GPT-6 Astra–RoboProbe configuration. 🤖 Physical-robot validation: Across 19 tasks, zero-shot Qwen scores 49.50/100, outscoring both π₀.₅ (47.75, fine-tuned on 10 hours of demonstrations per task) and Dream-1 (35.50, evaluated zero-shot). The agent uses depth; both baselines use RGB only. 🛠️ Unified for generalizable agentic training: Shared motion, perception, skill, and episode interfaces make trajectories across environments and robot embodiments easier to combine for future training. 🧠 What to train next: Analysis identifies visual understanding, spatial grounding, geometric coding, and embodied reasoning as joint priorities for future training.

NVIDIA NemotronLabs RE-USE is on ModelScope: a 9.6M-parameter speech restoration model for noise, reverb, clipping, codec artifacts, packet loss, and low-quality microphones. 🤖 🎧 Broad audio support: Handles mono audio from 8–48 kHz, with optional bandwidth extension and chunked inference for long recordings. 🏆 Strong restoration results: Leads the listed open alternatives on key quality, intelligibility, and fidelity metrics across the reported URGENT 2025 non-blind test subsets. 🗣️ Preserve the voice: Improves speech quality while retaining intelligibility and speaker characteristics, prioritizing the main speaker in overlapping speech. 📜 OpenMDW License Agreement v1.1; commercial use supported per the model card.

More languages, more natural delivery, still 0.6B. Audio8-TTS-Realistic-Air updates the Audio8 TTS stack with improved realism and zero-shot voice cloning. 🤖 🌍 23 supported languages: Adds 12 languages to Preview, including Arabic, Hindi, Portuguese, Russian, Thai, Turkish, and Vietnamese. 🎙️ Clone a voice or start from text: Use a clean reference recording with its exact transcript for zero-shot voice cloning, or generate speech without a reference. 🪶 Compact by design: The main DualAR model has 601M parameters, while the bundled 44.1 kHz codec handles both reference-audio encoding and waveform decoding. ⚡ Built for serving: The SGLang Omni adapter adds paged attention, dynamic batching, and an OpenAI-compatible /v1/audio/speech API. 📜 Apache-2.0 License. Quality varies by language and reference-audio quality, so review generated speech before final use.

Qwenputer explores a fascinating idea: embedding computational machinery directly into LLM weights, without retraining the original language model. Developed by @divagr1925 , the project has evolved from a RISC-V CPU inside Qwen3.5 into a family of models exploring new ways to extend LLM capabilities. The full model family and an interactive demo are now available on ModelScope: 🚀 Explore all models: 🧪 Try the interactive demo: Here's a closer look at the model family: • Qwenputer embeds a functional RISC-V CPU inside Qwen3.5, using Gated DeltaNet's recurrent state as RAM. • Qwenputer-Filesystem introduces persistent file storage within the model's recurrent state, enabling memory across sessions. • Qwenputer-NPU extends the internal CPU with neural network instructions, allowing it to execute small models. • QwenPU takes things further with a dedicated vector unit capable of running Karpathy's 260K-parameter Llama inside Qwen itself. Original project 👉:

English TTS in just 1.4M parameters, vocoder included. PiVo-1.4M runs offline on a CPU with NumPy and ONNX Runtime, without PyTorch. 🤖 🪶 1.4M parameters, vocoder included: Generates 16 kHz mono speech with a compact acoustic model and FARGAN-derived vocoder. 🖥️ Built for local, offline use: Generate with either English voice, Pi or Vo, directly on CPU and begin playback in 10 ms audio frames after text encoding. 🗣️ Words come through clearly: PiVo reaches 0.532% WER on its held-out test set. The authors note that this measures intelligibility, not general English quality or naturalness. 🔤 Predictable English pronunciation: It expands numbers, currencies, ordinals, and ISO dates, while unsupported words raise a clear error instead of silently guessing. 📜 Code and acoustic weights are Apache-2.0. The FARGAN-derived vocoder uses Xiph/Opus BSD terms, and the pronunciation dictionary retains its CMU license.

Scientific reasoning shaped by laboratory experience. Hydrogen-1.0 is an open-weight model from Deep Principle, built on GLM-5.3 with 1M context for chemistry, materials science, and life sciences. 🤖 🏆 Stronger scientific reasoning: Outperforms GPT-6 Astra on FrontierScience-Research in the authors' evaluation, while improving over GLM-5.3 on GPQA Diamond. 🔬 Learn from scientific work: Combines SFT and RL using domain knowledge, laboratory records, and complete discovery trajectories, including failed attempts and revised hypotheses. 🧬 Explore promising materials: After five search rounds, the mean best-so-far HOMO–LUMO gap reaches 4.318 eV, versus 3.689 for GPT-6 Astra and 3.353 for Kimi-K3 in the transition-metal-complex evaluation. 🛠️ Research with tools: Demonstrations span simulation debugging, reaction-network exploration, NMR analysis, and experimental protocol planning. 📜 Weights: MIT. Independently authored supporting code and documentation: Apache 2.0.

Introducing SenseNova-Loop-9B-Preview: a 9B open-weight multimodal model with 256K context for document understanding, coding, and agent workflows. 📜 Apache-2.0 License. 🤖 • Leads six of eight tasks in the team’s reported comparison: Scores 71.38 on BIRD-Dev, 66.89 on WorkBuddy Bench, and 58.87 on AIDABench. • A deeper path without duplicate weights: Its 24-layer middle block runs twice with shared weights, producing 56 layer executions from 32 distinct decoder layers. • Trained for long, multimodal work: Training recipes combine text, images, code, and tool-use trajectories at sequence lengths up to 262K tokens. • Acceleration and serving paths are available: MTP is supported now, with documented Transformers, LightLLM, SGLang, and vLLM deployment paths. Note: This is a Preview release; the technical report is still forthcoming. Examples created with Raccoon Work 👇

Beyond ordinary faces. Krea 2 OpenCast LoRA is an open-weight adapter by Akatz Labs based on Krea-2-Raw, creating fictional characters with exaggerated features, sculptural hair, and lifelike skin and clothing. 🤖 🎭 Design an original cast: Describe facial geometry, hairstyle, wardrobe, expression, and lighting to create fictional human or dressed animal characters. 🧪 Start from a proven recipe: Browse 18 exact prompts and seeds, then edit the included portrait workflow to make the character your own. 🛠️ Run it in ComfyUI: Use Krea 2 Turbo with the OpenCast LoRA at strength 1.0. No trigger token or custom nodes needed. 📜 Krea 2 Community License Agreement applies, including commercial-use conditions.

One model for points, boxes, masks, and alpha mattes. WeVisOne-VL-4B is open-weight, a visual perception model from Tencent's WeChat Vision, built on Qwen3-VL. 🤖 🎨 Try the demo: 🏆 Leading across 30+ perception benchmarks: The team reports strong results from object-level tasks to pixel-level understanding, including 71.3 on Dense200, 89.6 on PixMo-Count, and 71.2 on ReasonSeg. 🧩 One model, seven perception tasks: Detect, localize, count, point to, read, segment, and matte objects through structured next-token prediction. ✂️ From boxes to clean alpha mattes: A lightweight 9.6M pixel decoder produces masks and full-resolution continuous alpha mattes without relying on SAM. 🚫 Rejects targets that are not there: Reaches a 96.4% rejection rate on WeRef-Bench, helping avoid confident boxes for absent queried objects. 📦 Structured outputs for downstream workflows: Returns normalized boxes, points, masks, alpha mattes, and None when no matching target is found. 📜 Apache-2.0 License.

Sage-A1.5 is here, bringing 256K context, parallel research, and evidence checks to long-horizon scientific workflows. 🤖 🏆 SOTA across scientific benchmarks: Scores 90.0 on PinchBench, 81.0 on FrontierScience-Olympiad, and 48.3 on FrontierScience-Research, ahead of Kimi-K3, GLM-5.3, and GPT-5.6-sol in the reported comparison. 🧩 Agentic Orchestration: When enabled, delegates independent investigations to parallel subagents with isolated contexts, then integrates their evidence and findings to guide subsequent work. ✅ Verification: Optional Verification Mode reviews progress, checks evidence and assumptions, and redirects stalled searches within the model's own reasoning, without a separate critic model at inference. 🔬 Open Scientific Discovery: Supports exploratory research through tool-assisted analysis and produces inspectable research artifacts, extending its use beyond predefined question answering. 🛠️ Native tool use for research: Supports APIs, code interpreters, search engines, and external tools for tool-augmented scientific reasoning and professional workflows. ⚡ Available in BF16, FP8, and GGUF F16/Q8/Q4 formats, with Transformers, vLLM, and SGLang support. 📜 Apache 2.0.