Announcement recommending users upgrade from the MTP head to the DFlash 2 draft model for Qwen3.8-Flash-Next speculative decoding in llama.cpp, claiming ~17% average speedup, with Q4_0 quantization recommended as it matches Q8_0 acceptance/speed while using 287 MiB less memory. The card also details runtime memory trade-offs between the MTP head and the DFlash 2 drafter, noting the smaller on-disk drafter costs about the same at runtime due to its larger compute buffer.
LiquidAI 的 LFM3-Omni-2B 成为首个可在浏览器中通过 WebGPU 100% 本地运行的全模态模型,无需独立 ASR/TTS,在 M4 Max 上实现约 280 token/s、90ms 首 token 延迟。
Canva 推出了其首个视觉语言模型 Brush,能够理解设计的布局、图层与风格并协助用户编辑,被视为传统创意平台在 AI 时代自研模型的重要转折点。
Qwen4 has been released, with hints of a 120B parameter model surfacing via a GitHub PR in the qwen-code repository.
Unsloth has released a Dynamic V3.1 quantized GGUF version of DeepSeek-V4.1-Flash-1017, enabling the model to run locally on devices with just 64GB of RAM. It claims up to 3× faster inference, support for a 1-million-token context window, and across-the-board leadership in benchmarks against comparable quantized models.
A community discussion asking for real-world user experiences with Nanbeige4.3-33B, a 33B parameter open model featuring a 512K context window and strong benchmark results in reasoning, coding, and agentic tasks, despite concerns about high KV cache memory usage.
Unsloth 发布 DeepSeek-V4.1-Flash-1017 Dynamic V3.1 (preview) GGUF 量化模型,宣称在 7 项基准上精度保持优于 ISTA 的 GSQ-RCO GGUFs,并公开了无法复现对方 AIME25 100% 精度恢复声明的对比结果。
SVPino 介绍了一款名为 Mischief 1 的图像到视频生成模型,据称比其开源基础模型快 5 倍、推理成本低 80%,非常适合低成本快速迭代创意。
SenseNova releases SenseNova-Loop-9B-Preview, a looped multimodal model built on a shared looped block and mHC residual stream architecture, targeting long-context and agentic workflows with capabilities spanning motion design, PPT generation, and data analysis.
Gemma-4-31B-Humanlike-Chat 是基于 Google Gemma 4 31B 的微调模型,通过 SFT 与 on-policy 蒸馏训练,实现了无需提示的自然人类化文本风格,同时在 When2Call、BFCL 等工具调用基准上超越原版,且知识能力与基座持平。
inclusionAI 发布了 ZwZ2 系列(27B/9B/4B),基于 Qwen3.6 微调的视觉语言模型,通过 EvidenceConnect 构建的多区域验证数据与 GRPO 训练,在六大领域 25 个基准上均超越基座模型,总体准确率从 55.52 提升至 60.51(+4.99)。
HuggingModels 分享了一个基于 RoBERTa 的模型,可自动对漏洞报告的严重程度进行分类,训练数据来自真实的漏洞评分,旨在帮助安全团队更高效地处理海量 CVE。
NVIDIA 发布 Metatron v1,一个统一的 Unitree G1 人形机器人控制策略,将行走、场景交互与恢复整合为单一 MaskedMimic 风格 transformer 学生模型,并提供原生 PyTorch 检查点与 ONNX 浏览器导出。
NVIDIA releases AlpaJudge, a family of learned judge models for evaluating autonomous-driving reasoning, with reference-based and reference-free checkpoints available on Hugging Face under the OpenMDW-1.1 license.
Pika introduces H3 Flex, a post-trained MiniMax H3 text-to-video model that generates in six steps (vs. 50), supports up to 12 reference files, exports five audio stems, finishes at up to 4K, and is released with open weights — ranking #1 among open-weights models on DesignArena's leaderboard.
Fish Audio 为 Drama 3 语音模型发布了升级后的权重,使其在 Artificial Analysis 的 Provider Voice 和 Controlled Voice 两个 Speech Arena 榜单上均位列开源权重模型第一,并在发音鲁棒性基准上排名第一。
Figma 详细介绍了其正在训练的「Figma-fluent」编辑模型,该模型专注于帮助设计师在现有画布内容上进行间距调整、颜色修改、布局重构等迭代式、难以自动验证的日常编辑操作。
Arc-Transcriber is a ~70MB streaming speech model that performs ASR and NER/entity-linking in a single step, outputting transcripts, linked names, numbers, speaker roles, and audio events every 100ms via trained heads on a shared encoder, eliminating the usual multi-model pipeline. Two variants (Arc mini and Arc-v1) are released with weights on Hugging Face and code on GitHub.
Xenova 宣布突破:首个 100% 本地在浏览器中运行的真正全模态模型 LFM3-Omni-2B,借助自定义 WebGPU kernels 实现约 280 tok/s,支持视觉、音频和文本输入并生成交错的文本与音频输出。
Release of Swift 1.5 Flash Next - Heretic v2, a custom GGUF quantization of Qwen 3.8 Flash optimized for reverse engineering and coding, featuring new heretic weights and a layer-aware bit allocation strategy based on activation-energy auditing.