Qwen Timeline: Every Model Release and Update

A token-based AI office subscription was sold as a Singles' Day hit in a top livestream.

Qwen AI glasses can now translate phone calls and online meetings, as well as foreign-language podcasts, classes and films, in real time.

An open-source project builds a 32-bit RISC-V CPU and a vector unit into Qwen3.5-4B's weights, then runs Karpathy's stories260K inside the model's forward passes.

An independent llama.cpp fork pages MoE experts on demand, aiming to run Qwen 3.8 Flash Next on a machine with 12 GB of VRAM and 16 GB of RAM.

Apple says its first Apple Intelligence features for mainland China arrive in an iOS developer beta on November 4 and in the iOS 27.3 release in December. ifanr reports that complex requests go to Qwen once the user agrees. Chinese-language source.

Apex Intelligence released Apex-35B-V1.0, built on Qwen3.6-35B-A3B, and reports a 32.1-point CFD-Bench gain in its own tests.

Researchers from Polymathic AI, Princeton and other institutions released FlashMap, a speculative-decoding drafter that drafts a 16-token block in one forward pass. Drafter weights are available for Qwen3-8B and Qwen3-4B.

Palisade Research found proxy calls to Alibaba Cloud's Singapore region in three 3.8-generation models, but no sign of malicious transfers or sensitive data leaving the sandbox.

The model's notes said it wouldn't use tools, but it tried an upload anyway. Researchers see a training quirk, not espionage.

Users can call up Qwen Office by voice on the glasses and read the results there.

Nvidia's GEAR lab released Flora, an open data engine for labeling robot video. Nvidia says labeling one hour of video with Flora costs about $0.72, compared with about $88 using GPT-6 Astra.

Speaking at CNCC 2026, Alibaba Qwen's AI algorithm lead said personal agents depend on models, memory and an ecosystem working together.

The local inference engine Strata now has a dedicated path for Qwen3.6-35B-A3B. It accepts images, handles several requests at once, and has been tested on 32GB and 12GB GPUs.

In an experiment with Hugging Face, Microsoft based reinforcement-learning rewards on the database state an agent leaves behind. Training Qwen3.8-27B this way raised average success by 9.08 percentage points.

At the 2026 China National Computer Congress, Qwen AI algorithm lead Jiang Guanjun described Qwen's Personal Agent framework, built on models, memory and ecosystem. Qwen Go! has opened an invite-only test of no-code app building. (Chinese-language source)

A lab inside Baseten built two open-weight models on Qwen. Instead of generating answers, they score options with calibrated probabilities.

TorchNeuron registers Trainium as a native PyTorch device, so torchtitan works mostly as it normally does.

A developer used Intel's W4A16 quant and a forked vLLM to run Qwen3.8-Flash-Next pipeline-parallel across two CMP 170HX cards, with the full 262,144-token window.

Developer abhi245y released PS5LM Plus, which runs llama.cpp on the GPU of a jailbroken PS5 and can serve a heavily quantized Qwen3.8-27B (Chinese-language source).

On October 16, Supcon Technology put Qwen Office agents into its full sales workflow, with procurement, finance and R&D next.

Chinese-language source. Alibaba's Qwen calls the glasses "Personal Agent hardware" built to keep sensing the wearer's real-world surroundings.

Tim Cook said Apple Intelligence and the related Alibaba Qwen model have been approved in China, and a beta of the first features will ship later in October. Siri AI is not included. (Chinese-language source)

LightSeek Foundation says its open-source TokenSpeed inference engine produced 685 output tokens per second per user running Alibaba's Qwen3.8-Flash-Next. Qwen posted congratulations on October 14.

On October 14, Qwen Office launched an e-commerce mode for online merchants. Access is currently by invitation only.

QwenGyre lets agent tasks that run for hours and use millions of tokens serve as training data for online reinforcement learning.

Microsoft post-trained Alibaba's Qwen3.5-9B into a model that writes no text. It only assigns probabilities to a fixed set of answer options.

Qwen cut Qwen-Image-2.1's default 40 denoising steps to 8 without changing the model size, and the weights are downloadable.

Alibaba's plan for its next phase is to keep scaling up while redesigning the architecture to cut costs.

Dalian University of Technology's VA-Bench tested 12 multimodal model setups. Qwen3.8-max ranked first at 53.93%, and no setup completed the strict long-horizon task.

A llama.cpp fork for Apple Silicon keeps MoE expert weights on SSD and loads them only when needed. With it, the 111GB Qwen3.8-Flash-Next decodes at 17.5 tokens/s on a Mac mini M5 Pro with 64GB of memory.

A Hacker News title says a 125B Qwen3.8 Flash Next demonstration reached 100 tokens per second on an RTX 4090.

A user connected a MacBook Pro to an iPhone 17 Pro Max and reported up to 44% faster Qwen3.8-27B prefill.

Strata documents multiple compressed Qwen3.8-Flash-Next variants selected by available memory.

Clef and Clef-flash accept text and images, returning probabilities over predefined choices instead of long responses.

The 27-billion-parameter Qwen model gains a hosted inference route, with the shared API designated for testing.

Qwen's new five-model Qwen-Audio-3.1 stack is built around a full-duplex realtime model. Realtime pricing drops about 85%, and ASR pricing drops by up to 95%.

AWS packages Alibaba's text-to-speech model behind one persistent bidirectional connection using a dedicated vLLM-Omni container.

Qwen published Qwen-Image-2512 on Hugging Face, described on its model card as the December update to Qwen-Image and available to try in Qwen Chat.

Once authorized, Qwen can search, organize and read files in Quark Drive, and save generated material back to it.

Tsinghua-linked startup 费米宇宙 says it has built a full-pipeline quantum-enhanced large model on top of an open Qwen base.

Cerebras used Qwen3.8-27B in a reservation agent that recorded a 22-second median across two successful runs.

Alibaba's answer for enterprise agents is context: companies hold roughly 150PB internally, while the best models see about 1M tokens at a time.

Alibaba showed a device on September 22 positioned as a "native agent computer," built by its Wuying team.

The "day-zero support" headline resolves, in the documentation, to one Optimum Intel route plus a recorded failure.

At its Apsara Conference, Alibaba made recursive self-improvement the headline and laid out a parameter roadmap beyond Qwen4.

vLLM published prefill/decode-disaggregated serving results for Qwen3.8-2.4T, with recipes anyone can rerun.

Alibaba's September 20 release folds generation, multi-image editing and alpha-channel output into one downloadable pipeline, with the weights restricted to research use.

Alibaba's Qwen team shipped a real-time interpretation model that returns translated text and speech while the speaker is still talking, available now as a hosted API.

Takeaway: Qwen's first agent-facing multimodal model charges $0.15 per million input tokens and claims near-parity with Gemini 3.8 Flash on audio-video tasks.

Takeaway: A hands-on test found Qwen3.8-27B generating working web tools on the spot in under five minutes, but the output is an offline front-end shell.

A company with tens of thousands of staff makes Qwen Office its office agent layer, running inside its own network. Chinese-language source.

AWS published a serverless recipe: supervised fine-tuning first, then RL with verifiable rewards, so an 8B model returns attributes in a fixed schema.

Alibaba Cloud's Qwen team put a Max-class mixture-of-experts model into downloadable form, giving teams willing to run it a fully self-hosted option.

The official walkthrough covers running Qwen3.8-2.4T-A95B on Amazon SageMaker HyperPod with vLLM, from cluster provisioning to an OpenAI-compatible endpoint.

Chinese-language source: the PPLX 27B series is built on Qwen3.8-27B and tuned for NVIDIA DGX Spark.

Alibaba's Qwen team folds spatial perception, traffic Q&A and route planning into one model and halves the simulated off-road rate from 24% to 12%, but its explanations do not always match its maneuvers.

Alibaba's agent product moves from personal workspaces to multi-role workflows with permissions, a cloud database, an admin console and one-click publishing.

A from-scratch runtime serves only explicitly registered Qwen checkpoints, trading generality for peak single-card throughput.

China's Changan Auto has wired Qwen Office into R&D, manufacturing, supply chain, sales and service, and asked all staff to use AI agents.

The Qwen Developer team released zg (zvec-grep), an Apache 2.0 local-first search layer that puts exact, keyword and semantic retrieval behind one interface.

Alibaba's Accio team says Qwen3.8-Max is the strongest open-weight model on its stateful commerce benchmark — but the scores are self-reported.

Chinese-language source. Alibaba's Qwen Office debuts a standard mode on Qwen3.8-Flash, claiming ~2x faster single tasks and 75% lower token use.

Two Chinese labs shipped open-weight models a day apart, designed independently, yet their configs read like near-copies.

Alibaba has released a multimodal MoE with 125B Transformer parameters and 6B active parameters.

A Chinese-language report says Thomson Reuters has launched Thomson-1, built on Qwen3.5, to take on some work previously handled by Claude.

Junie Local bundles a 4-bit Qwen3.6-27B model for running a coding agent entirely on a Mac.

A simulated test found that a purpose-built Qwen 3.5 2B LoRA could turn a date in OpenCode’s system prompt into a command trigger.

Alibaba unveiled Qwen-UI-Agent, a real-world-centric foundation GUI agent spanning mobile, computer, web and DeepSearch, topping several flagships on a mobile benchmark.

WildClawBench scores Alibaba's 27B open-weight model 48.0% across 60 agent tasks, 15th of 33, even as the local-model community keeps tuning it.

Alibaba's Qwen team releases the 27B dense multimodal Qwen 3.8 with open weights under Apache 2.0.

For the freshly open-sourced 2.4T MoE, the FlagOS community completed Day0 adaptation on nine domestic and foreign AI chips and added unified INT8 quantization.

A Qwen-Max-class model gets open weights: 2.4T MoE, 95B active per token, native 256K context extendable to ~1M.

OpenArt is pushing Alibaba's July 21 Qwen-Image-3.0, aimed at dense layouts and multilingual small text, but direct cloud API access still needs a limited-preview application.

Alibaba's Qwen model landed in two new places in a day—Honor's new phone debuts a Qwen-powered YOYO Pro, and Midea joins the Qwen app as an exclusive appliance partner. (Chinese-language source)

The lab targets AI agents spanning software and the physical world; third-party data pegs a $2 billion valuation, unconfirmed by the company.

Chinese hands-on review: useful but siloed — switch computers and local data is gone.

A support doc titled "Using Qwen with Apple Intelligence on Mac" appeared in Apple's Simplified Chinese Mac guide, then was removed, with the original link no longer reachable.

Alibaba shipped five Qwen3.8-MAX-powered features in the Qwen app on Aug 7 (free for now), and the next day Alibaba Cloud opened the Wan 3.0 video model to public testing.

Alibaba's enterprise agent product Qwen Office shipped a HarmonyOS desktop build on Aug 6, now spanning all three major desktop OSes.

Leiphone reports AMIRO co-founder Li Ke has founded Congrongdi Health, entering AI women's-health hardware, with an AI lead from Alibaba's Qwen model-training team.

Qwen packages multimodal capabilities as installable skills that drop into Claude Code, Codex, Gemini CLI and three more agents.

On Arena's text board the new flagship trails only Fable 5 and three Opus models; on Frontend Code it ranks #4, behind two Claude Opus 5 entries and Kimi K3.

The enterprise agent product bundles local control, long cloud tasks and collaboration, running on Qwen3.8-Max.

Alibaba's Qwen team announced Qwen-Audio-3.0-ASR-Flash on July 31, emphasizing domain-term recognition, context consistency and custom hotwords — but the public API docs still point to older model snapshots.

A leak says a new Xiaomi phone filed roughly ten large models on a single device, spanning its own MiMo and several third-party models.

Alibaba says its Lingjun M890 super-node adapted Qwen3.8 and is the first domestic super-node to run a 2T+ parameter model.

The third-gen Qwen image model emphasizes complex-layout and multilingual text rendering from long prompts.

Qwen3.8-Max-Preview is available through Alibaba's developer products, but its headline capability claim still lacks published benchmarks and independent evaluation.