Anthropic’s September threat report says distillation campaigns tied to Alibaba, Moonshot AI and DeepSeek pulled more than 186 million exchanges from Claude, and that two labs quietly served Claude to their own users. We read the source: seven labs, five counts, windows from 14 days to three months, a June Senate letter that does not reconcile, and two Qwen releases that predate both counts.
TestingCatalog has surfaced Hybrid mode, an unreleased setting in Perplexity Computer’s Mac app that keeps orchestration in the cloud while delegating suitable subtasks to a model running locally, plus a Privacy Gate model that checks outbound data for personal information and asks the user before it is sent. This article covers the three downloadable local models (a 19GB Perplexity model, Qwen 32B at 17.4GB, and a 5.6GB Gemma-based model for 16GB Macs), which Mac mini configurations qualify, how Privacy Gate differs from June’s hybrid agentic inference design, the credit arithmetic the feature targets, how it compares with the Nvidia-only Portable Computer, and what a UK business should do about a feature that is still hidden and undated.
Accio, the commerce agent team inside Alibaba International, has open-sourced CommerceAgentBench: 107 long-horizon business tasks running against fourteen offline replicas of real software, each in a fresh container, each graded on the state the agent actually changed. This breakdown covers the task mix, the grading design, the three-harness leaderboard led by Claude Opus 5 at 66/107, the twelve-task swing that a harness alone can produce, the caveats the README discloses about managed endpoints and unmatched reasoning effort, the two-licence split, the sibling Business Arena benchmark, and what any of it should change if you are buying agents this quarter.
Alibaba open-weighted Qwen3.8-Flash-Next on 26 August 2026: a multimodal mixture-of-experts model with 125 billion parameters, a separate 51-billion-parameter N-gram embedding table, and just 6 billion parameters activated per token. This breakdown covers the four rebuilt subsystems — Gated DeltaNet paired with Qwen Sparse Attention at block granularity, a Gated Residual stream widened to four gated branches, the N-gram table that offloads to host RAM, and the Muon plus AdamW training recipe with batch-size warmup removed — alongside the 48-layer stack of 512 experts that fires eleven per token, the published benchmark table showing 62.5 on SWE-bench Pro against 53.4 for Claude Opus 4.6 and 84.5 on AndroidWorld against 62.0, the single loss on Humanity’s Last Exam at 35.9 against 40.0, the unverifiable one-ninth training cost claim, the 262,144-token native context extended to a million with YaRN, hosted pricing of $0.16 and $0.47 per million tokens against $2.00 and $6.00 for Qwen3.8-Max, the real hardware bill from a 172.78 GiB FP8 checkpoint down to a 111 GB four-bit GGUF, the qwen-community-1.0 licence that is not Apache 2.0, and a buyer’s checklist for treating a preview checkpoint as a production dependency.
Perplexity announced Portable Computer on 25 August 2026, built in close partnership with Nvidia: the full agentic Computer stack — model, inference engine, agent harness, tools, connectors and sandbox — running on your own hardware, where local work consumes no billing credits. It needs an Nvidia RTX GPU with at least 24GB of VRAM or a DGX Spark, runs on DGX OS or Ubuntu with Windows due in September 2026, and offers Qwen 3.8 27B, Perplexity’s post-trained PPLX 27B and, soon, Nvidia’s Nemotron 3.5 Lightning. This breakdown covers the hardware floor, the per-step cloud escalation flow, Perplexity’s own benchmark numbers, what zero token costs actually excludes, and who should pilot it now.
Qwen3.8 27B goes open-weight at 00:00 JST on 15 August 2026. This launch-day guide separates what Alibaba has confirmed from what is still unpublished — above all the licence, after the sibling Max shipped under bespoke terms. It maps VRAM and quantisation tiers from a 24GB card to an H100, and sets the Qwen3.6-27B baseline of 77.2 SWE-bench Verified as the bar the new weights must clear.
Qwen3.5 Omni combines audio, video, image, and text handling with speech output inside Alibaba Cloud’s OpenAI-compatible workflow.
A highly SEO-optimized news explainer on Qwen3.6-35B, the open-weight Qwen3.6-35B-A3B release, covering agentic coding, multimodal support, long context, and deployment details.
