Every story across the Writeble desks, newest first.
On September 15, 2026, researchers announced ShadowPEFT as a first-class method in Hugging Face’s PEFT library—a stateful shadow-network adapter with the usual get_peft_model API, available now from PEFT main ahead of the next release.
Anthropic’s September 2026 multi-lab threat report details illicit Claude distillation it says it disrupted across seven China-based labs — a broader peg than Writeble’s June 28 Alibaba-only piece — with TechCrunch and The Hacker News relaying Anthropic-reported volumes and caveats.
UC Berkeley’s RDI introduced CUA-Lite in September 2026 as an open platform that unifies computer-use agent environments, supervised data, and model harnesses across desktop, browser, and mobile — with Lite.Gym, Lite.Sample, and 15+ plugged-in benchmarks.
Shanghai AI Laboratory’s InternLM org published Atria Dawn Preview on Hugging Face around September 11–12, 2026—MIT-licensed agentic MoE weights with an FP8 sibling, a 256K context listing on the model card, and vendor-reported agentic benchmarks, without a formal launch post.
Amazon engineers released Pizza Bot as a standalone Apache 2.0 community project that gives long-running AI agents an email-style inbox, with DeepAgents and LangGraph checkpoints so work survives disconnects and approvals.
Wiz Research found 294 of 3,074 internet-facing LiteLLM gateways still accepted the docs example admin key sk-1234, a credential that can expose provider API keys—while related LiteLLM MCP flaws are already on CISA’s KEV list.
Tencent’s Hunyuan Speech Team and academic partners released Gander, a 9B Apache 2.0 omni-interaction model for streaming speech, vision, and async agent tasks, with weights on Hugging Face and an open runtime on GitHub.
DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026 as a 552B MoE multimodal open-weight model under MIT, and starting 04:00 UTC on September 14 its API routes deepseek-v4-pro requests to Flash at Flash rates until V4.1-Pro ships.
Google Research’s ToolGrad framework builds verified tool-use chains before writing user queries, shipping Apache 2.0 code plus the ToolGrad-500 dataset and Gemma-3 fine-tunes that the authors report near frontier BFCL scores.
Cohere and Cohere Labs released North Small Translate on September 10, 2026—a 218B-parameter MoE translation model with about 25B active parameters—under CC BY-NC 4.0 on Hugging Face, with vendor-reported WMT26 scores led by an 83.6 all-languages figure.
OpenAI opened the Agents API in public beta on September 10, 2026, exposing the managed Codex harness for durable cloud agents with hosted, self-hosted, or partner sandboxes—no separate API fee beyond model, tool, and container usage.
Ant Group’s inclusionAI lab released Ling-3.0-flash-VL under MIT—a 124B MoE multimodal model with about 5.5B active parameters—with BF16 weights on September 4 and FP8 on September 8; BlockBeats framed the open-source release on September 9.
Google’s Agent Development Kit for Kotlin reached 1.0 GA on September 9, 2026, with full ADK Core parity plus Android-first extensions for LiteRT-LM, ML Kit, Firebase AI Logic, Room sessions, and AppSearch memory under Apache 2.0.
IBM released Granite Time Series PatchTST-FM-r2 on September 9, 2026—a ~385M-parameter zero-shot forecaster dual-licensed under Apache 2.0 and OpenMDW 1.0, ranked among top replicable zero-shot models on GIFT-Eval as of September 8.
OX Security disclosed CVE-2026-82533 (CVSS 9.4) in DeepSeek Harness, where an unauthenticated local control API trusted via Host header lets a sandboxed agent elevate to danger-full-access; fixed in 0.1.2-alpha.1.
GitHub’s Project HydraFusion research preview in Copilot CLI builds single, cascade, or critique workflows across models, cutting estimated cost versus Claude Opus 5 on three benchmarks while matching or beating quality on only one of them.
Nex-AGI published Apache-2.0 weights for Nex-N2.5 Mini (~35B/~3B active multimodal) on Hugging Face September 8, 2026, with a 2×H100 SGLang recipe; Pro multimodal weights remain coming soon.
NVIDIA Personal AI Router (PAIR) is an Apache 2.0 virtual inference router that discovers home-network machines and spreads independent Ollama or LM Studio requests across them, cutting a five-subagent Hermes demo from 18 minutes to 8:48 on a three-device cluster.
NVIDIA opened CUDA Rust with two NVlabs projects—cuda-oxide for SIMT kernels and cutile-rs for Tile kernels—so Rust can compile natively to PTX and catch buffer aliasing at compile time, with cutile-rs already used in Hugging Face Grout and mistral.rs.
Thesys released OUI-1, a DiffusionGemma 26B-A4B finetune that writes openui-lang interfaces, scoring 71.7% on the Generative UI Benchmark versus 13.0% for the base model, under Gemma Terms of Use.
Qwen and Huazhong University of Science and Technology released Qwen-Drive-1.0-4B, an Apache 2.0 vision-language driving model that unifies 3D perception, traffic VQA, and motion planning on a shared Qwen3.5-4B backbone.
