ModelScope

Z.ai just fixed critical Coding Agent serving bugs on GLM-5. 🔧 Garbled outputs, repetition, rare-character generation — only surfacing under high-concurrency, long-context workloads. Weeks of debugging, here's what they found and fixed👇 Worth a read if you're running Coding Agent infrastructure at scale. z.ai/blog/scaling-p… Bug #1: KV Cache race condition under PD disaggregation — stale RDMA writes corrupting cache of new requests. Fixed by enforcing explicit sync between request termination and KV Cache write completion. Anomaly rate dropped from ~0.1% to below 0.03%. Bug #2: Missing load-use ordering in HiCache — Forward Stream reading KV Cache before Load Stream finished. Fixed by inserting explicit sync points before Indexer kernel launch. Already submitted to SGLang as PR #22811. 🌟 Bonus optimization: LayerSplit — layer-wise KV Cache partitioning to reduce per-GPU memory footprint. Up to 132% throughput improvement on long-context Coding Agent workloads. Free API inference is also available on ModelScope — try GLM-5 today: modelscope.cn/models/ZhipuAI…

@Zai_org just fixed critical Coding Agent serving bugs on GLM-5. 🔧 Garbled outputs, repetition, rare-character generation — only surfacing under high-concurrency, long-context workloads. Weeks of debugging, here's what they found and fixed: Bug #1: KV Cache race condition under PD disaggregation — stale RDMA writes corrupting cache of new requests. Fixed by enforcing explicit sync between request termination and KV Cache write completion. Anomaly rate dropped from ~0.1% to below 0.03%. Bug #2: Missing load-use ordering in HiCache — Forward Stream reading KV Cache before Load Stream finished. Fixed by inserting explicit sync points before Indexer kernel launch. Already submitted to SGLang as PR #22811. 🌟Bonus optimization: LayerSplit — layer-wise KV Cache partitioning to reduce per-GPU memory footprint. Up to 132% throughput improvement on long-context Coding Agent workloads. Worth a read if you're running Coding Agent infrastructure at scale. z.ai/blog/scaling-p… Free API inference is also available on ModelScope — try GLM-5 today 👇 modelscope.cn/models/ZhipuAI…

@natolambert It’s nice to meet you! 🤣

🤗

🤖

So excited to announce Ling-2.6-1T is now live on ModelScope!🔥 This 1T parameter model is built for complex agent workflows, multi-step execution and long-context understanding. It truly delivers in production. 📊The benchmarks speak for themselves: - AIME26 — leads all non-reasoning models - SWE-bench Verified, TAU2-Bench, BFCL-V4, PinchBench — first-tier open-source - ~16M tokens on Artificial Analysis full eval — same efficiency story as Ling-2.6-flash Works with Claude Code, OpenClaw, OpenCode & CodeBuddy ✅ SGLang & vLLM ready · Open weights available now 🚀 Explore on ModelScope 👇 modelscope.cn/models/inclusi… modelscope.ai/models/inclusi…

Updated Model Link: modelscope.cn/collections/An… Android Demo: modelscope.cn/models/AngelSl…

Tencent Hunyuan just open-sourced Hy-MT1.5-1.8B-1.25bit, a 440MB offline translation model for mobile. 33 languages, no internet required. 🚀 The 1.8B base model matches commercial translation APIs and 235B-scale models on quality benchmarks, 🔥 outperforming Google Translate and Baidu Translate. The 440MB version gets there via Sherry (1.25-bit sparse ternary quantization, accepted at ACL 2026): 4 params per group, 3 stored at 1-bit, 1 zeroed out, averaging 1.25-bit per param. Also available: 574MB 2-bit variant (SEQ quantization, near-lossless quality, faster on Arm SME2 devices). Android demo available now. 🤖 modelscope.cn/collections/An… 📄 modelscope.cn/papers/2601.07… 📄 modelscope.cn/papers/2512.24…

@KuittinenPetri They're going to update some information based on the team's feedback.

@ModelScope2022 Hey @ModelScope2022 ! The first link ( modelscope.cn/collections/An… ) is not working.

Ling-2.6-flash is now live on ModelScope.🚀104B params / 7.4B active. And it's built around one idea: every token should count. For real-world Agent deployment, that efficiency gap is everything. 💪 - 340 tokens/s inference speed on 4× H20 - Only 15M tokens to complete Artificial Analysis full eval — ~1/10 of Nemotron-3-Super - BFCL-V4, TAU2-bench, SWE-bench Verified — SOTA-level among same-size models Explore it 👇 modelscope.cn/collections/in… modelscope.ai/models/inclusi… Works with Claude Code, Qwen Code, OpenClaw & KiloCode ✅ FP8 & INT4 variants available now.

🌟We're happy to share SenseNova-U1 on ModelScope, a native multimodal model that unifies understanding and generation in one place. 📄Apache 2.0. Built on NEO-Unify with no visual encoder and no VAE, language and vision are treated as one unified composite end-to-end. One model covers T2I, editing, interleaved generation, and visual QA. 🏆 SOTA on open-source understanding and generation benchmarks. 8B-MoT (dense) + A3B-MoT (MoE). Try it now 🤖 modelscope.cn/models/SenseNo…

Xiaomi MiMo-V2.5 is now live!🚀a major step forward in agentic capability and multimodal understanding. 310B params / 15B active. Native vision + audio. Up to 1M token context. And it's open-source. ⚡ 🤖 Explore on ModelScope 👇 modelscope.cn/collections/Xi… The numbers back it up 📊 ✅Agentic: Claw-Eval 62.3 | SWE-bench Pro 56.1 | Terminal-Bench 2.0 65.8 ✅Multimodal: MMMU-Pro 88.5 | Video-MME 87.7 | DailyOmni 64.0 Matches Kimi K2.6 on agentic, Gemini 3 Pro on video Frontier-level agentic and multimodal in one unified model, at roughly half the cost of MiMo-V2.5-Pro.

✨ Inclusion AI's LLaDA2.0-Uni is open-source! A single MoE-based diffusion LLM that unifies visual understanding and image generation — natively, in one model. Download on ModelScope 👉 modelscope.ai/models/inclusi… Built on a single Mask Token Prediction paradigm, LLaDA2.0-Uni handles: 🖼️ Text-to-image synthesis at 1024×1024, with the option to "think" before drawing 🔍 Visual question answering, captioning, and document understanding on par with dedicated VLMs ✏️ Instruction-driven image editing — single or multi-reference, faithful to original details 🎨 Interleaved text-image reasoning, opening the door to a new class of multimodal chains Released under Apache 2.0 — paper, code, and weights all open. 📄 modelscope.ai/papers/2604.20… 🔗 github.com/inclusionAI/LL…

🎨 Qwen-Image-2.0-Pro demo is now live on ModelScope! 👉 Try the demo: modelscope.ai/studios/Qwen/Q… modelscope.cn/studios/Qwen/Q… Get a hands-on preview of its upgraded image quality, multilingual text rendering, and improved instruction following.

New on ModelScope! Meet MiMo-V2.5-ASR, an end-to-end speech recognition model, SOTA across a wide range of public benchmarks. 🚀 Download 👉modelscope.ai/models/XiaomiM… Specifically built for real-world scenarios: 🗣️ Chinese dialects: Wu, Cantonese, Hokkien, Sichuanese and more 🔀 Code-switching: Chinese-English, no language tags needed 🎵 Song lyrics: Chinese and English, mixed accompaniment 🔊 Noisy environments: heavy noise, far-field, adverse acoustic conditions 👥 Multi-speaker: overlapping conversations and meetings 📚 Knowledge-intensive: classical poetry, technical terms, proper nouns 📝 Native punctuation from prosody — no post-processing needed

🚀 DeepSeek V4 just landed! Explore the full family of DeepSeek-V4 on ModelScope today: modelscope.cn/collections/de… 4 open-weight models, all with native 1M-token context, including: 🔹 DeepSeek-V4-Pro: 1.6T total, 49B active, 1M context — frontier flagship 🔹 DeepSeek-V4-Flash: 284B total, 13B active, 1M context — speed-optimized 🔹 DeepSeek-V4-Pro-Base: pre-trained 1.6T foundation for frontier-scale post-training & research 🔹 DeepSeek-V4-Flash-Base: pre-trained 284B foundation for efficient domain adaptation 🏅Three inference modes — Non-Think / Think High / Think Max — let you dial reasoning effort on demand. In Think Max, V4-Pro reaches 93.5 on LiveCodeBench, 3206 on Codeforces, and 95.2 on HMMT 2026, closing the gap with leading closed-source frontier models on reasoning and agent tasks. 📄MIT licensed. #ModelScope #DeepSeek #DeepSeekV4 #OpenSource

Tencent HY just released Hy3 preview 👉open source. 295B total, 21B active, 256K context. Hybrid fast-slow thinking MoE. 🚀 First model after a full rebuild of pretraining and RL infrastructure. Biggest gains in coding and agentic tasks. 🛠️ Agent: drives up to 495-step complex workflows in production (docs, data analysis, MCP tool chains) ⚡ Inference: TTFT -54%, end-to-end latency -47%, success rate 99.99%+ on CodeBuddy/WorkBuddy 🎯 Strong on SWE-bench Verified, Terminal-Bench 2.0, BrowseComp, WideSearch — competitive across coding and search agent benchmarks ✅ OpenClaw / OpenCode / KiloCode compatible. vLLM + SGLang supported. 🤖 modelscope.cn/models/Tencent… 💻 github.com/Tencent-Hunyua…

@xeophon 👏👏👏

@ModelScope2022 congrats!!

Qwen3.6-27B is now live on ModelScope. 🎉 27B dense params. Flagship-level agentic coding. No MoE routing required. It just beat Qwen3.5-397B-A17B across every major coding benchmark 🔥 Download weights👇 modelscope.ai/collections/Qw… With the release of Qwen3.6-27B, ModelScope is rolling out training service for the model on Day-0, accessible via Tinker-compatible training APIs, with support for SFT, DPO, GRPO, and more. Powered by Twinkle✨, our open-source framework that enables client-server training paradigm. GitHub: github.com/modelscope/twi… Reference: modelscope.cn/organization/t…

@junyao_gao62882 congrats on the amazing job!

Many thanks to @ModelScope2022 for featuring MegaStyle and for the support! We’re excited to see MegaStyle shared with more people. Welcome to check out our code here: github.com/Tencent/MegaSt…