--- title: Nine teams put small Qwen models to work in one day description: Nine teams used SIE and Alibaba Cloud Model Studio to build routers, meeting tools, voice agents and image workflows at a one-day London hackathon. canonical_url: https://superlinked.com/blog/qwen-hackathon-london last_updated: 2026-08-25 --- On 14 August, we hosted a one-day hackathon with Qwen and Alibaba Cloud in London. Nine teams built with SIE and Qwen models on Alibaba Cloud Model Studio. By the evening, all nine had working demos: a self-updating model router, a financial news channel, local meeting tools and a voice assistant running on a Raspberry Pi Zero 2 W. ![Builders at work in the Alibaba Cloud London office during the Superlinked x Qwen hackathon](/images/blog/qwen-hackathon/hackathon-room-wide.jpg) ## Alibaba Cloud hosted the day, and Qwen engineers helped teams build Alibaba Cloud hosted the hackathon in its London office and gave every team Model Studio credits. Qwen engineers stayed in the room during the build, helping teams choose models and debug their integrations. ![The Alibaba Cloud and Qwen team introducing the day alongside Superlinked](/images/blog/qwen-hackathon/alibaba-cloud-qwen-team.jpg) ## Several teams split routine work from expensive calls A few teams independently divided the work between smaller models on SIE and hosted models on Model Studio. Darwin Router and TripWire made the split easiest to see. One routed by confidence; the other escalated when its security scanners disagreed. Echo used SIE for notes and search while Model Studio handled diarized transcription and cross-meeting synthesis. The implementations were different, but the reason was practical: use the larger model only where it changes the result. ## Darwin Router updated its routing examples as it ran **Kannappan Sirchabesan, [@hashkanna](https://github.com/hashkanna/darwin-router). Winner, Superlinked track.** Darwin Router embeds each request with `Qwen3-Embedding-4B` on SIE, then compares it with a vector index of labelled examples. Routine queries go to a small local model; harder ones go to `qwen-max`. When confidence is low, the router sends the same request to both models, asks `qwen-max` to judge the answers without knowing which model produced them, then stores the result as a new routing example. Kannappan started on Friday morning with an empty index. By the afternoon, the router measured 82.5% accuracy on a holdout set and an estimated cost roughly 60% below sending every request to a frontier model. It reached that point without a training run or fine-tune. ## AliBC turned its launch into the first news story **Dimitrios Koutsoumpos, [@dimknaf](https://github.com/dimknaf/AliBC). Winner, Alibaba visual track.** AliBC is an autonomous financial news channel. It writes the segments, generates the presenter and runs the ticker without a studio crew. ![AliBC broadcast still, an autonomous news anchor presenting the channel's own launch](/images/blog/qwen-hackathon/alibc-broadcast.jpg) Its first bulletin covered AliBC going live in London. The strap underneath read: "we don't sleep, we don't blink, we don't take lunch." ## Echo split meeting work between SIE and Model Studio **Mykyta Yakivets and Oriol Morros Vilaseca, [@yakivets](https://github.com/yakivets/superlinked_hack). Winner, social media prize.** Echo is a small meeting recorder that streams audio to its backend. Alibaba Cloud Model Studio handles speaker-diarized transcription and cross-meeting synthesis. SIE writes the notes, extracts entities, embeds each meeting and reranks search results. ![Echo meeting hardware on the table during the hackathon](/images/blog/qwen-hackathon/mythic-aida-hardware.jpg) ## Aida ran a voice assistant on a Raspberry Pi Zero 2 W **Dave Cheng, [@MasteraSnackin](https://github.com/MasteraSnackin/Aida-Relay). Mythic.** Aida runs on a Raspberry Pi Zero 2 W with 512 MB of RAM. It calls Qwen through Model Studio with SIE in the loop. The official Whisplay daemon has sole control of the screen, LED, button and audio, which prevents different parts of the assistant from fighting over the hardware. Aida adds selectable assistant modes, memory shared across sessions, games and an offline fallback for network outages. ![The Aida Relay device running in the palm of a hand next to superlinked.com](/images/blog/qwen-hackathon/mythic-whisplay-device.jpg) The complete device cost about £15. ## Sieve sent one of 20 ideas to Wan for rendering **Ritik Sah and Tanvir Farhad, [@ShieldTech-Ltd](https://github.com/ShieldTech-Ltd/sieve). ShieldTech.** Generating text candidates is cheap compared with rendering the final image and video. Sieve creates 20 directions for a campaign brief, then ranks and refines them through SIE using Qwen. It sends one candidate to Alibaba Wan 2.6 for the final render. On the demo brief, that meant one premium render call instead of 20. Sieve also records every candidate, its score, the model that reviewed it, the latency and the estimated saving. You can see why the winning direction survived and what the other 19 would have cost. ## TripWire escalated when its security scanners disagreed **Mani Sarkar, [@neomatrix369](https://github.com/neomatrix369/tripwire). TheSandbox.** TripWire scans AI skills and MCP servers with three scanners, then reviews their findings. `Qwen3.5-4B` on SIE reads every result. It calls `qwen-max` when the scanners contradict one another, coverage is incomplete or the evidence does not support a clear severity. ![TripWire dashboard showing risk-scored MCP servers and the SIE to Model Studio routing decision](/images/blog/qwen-hackathon/tripwire-dashboard.png) The dashboard records the full chain for each item, including the arbitration note. In one test, TripWire returned amber instead of red for a tool-poisoning finding because two of five scanners were unreachable and the completed scans did not yet support a red severity. ## RentalFinder simulated a landlord call in the browser **Richard Lao, [@richard7ao](https://github.com/richard7ao/hackathon_qwen). Octuple.** ![RentalFinder AI landing page, find your next rental without the chase](/images/blog/qwen-hackathon/rentalfinder-landing.png) Describe the flat you want in plain English. RentalFinder matches the request against live London listings, then generates a booking conversation for the chosen property. SIE runs the chatbot on Qwen 3.5 and matches listings with Qwen3 embeddings. The voice demo builds a scripted exchange between the renter's agent and the landlord. Qwen3-TTS generates audio for the agent's lines, which play in the browser; the landlord's replies appear as text. Qwen-Image produced the landing-page art. ## Venue Visualiser previews an event inside a listing photo **Rob Chu, [@RobChooses](https://github.com/RobChooses/venue-visualiser).** Rob built a Chrome extension for venue booking sites. Pick a photo of a space and choose an event setup; the extension asks the Qwen image model through Model Studio to generate a preview. It builds the prompt from a local template by default and can optionally use a configured text or vision LLM. Model Studio performs the image edit. ## Yappy keeps meeting audio and notes on the laptop **Lakshmikanth Karur Rajamani, [@krlakshmikanth](https://github.com/krlakshmikanth/yappy).** Yappy is a meeting copilot that runs on SIE on the user's laptop. The audio and notes stay on the machine, which makes it suitable for calls where joining a recording bot is not allowed or welcome. ## Small models handled most of the work By the evening, all nine teams had a demo. None had trained or fine-tuned a model, and none needed to provision an accelerator cluster. Where teams used both SIE and Model Studio, they were explicit about which work stayed local and which requests moved to a hosted model. The dividing line was different in each project: confidence, privacy, rendering cost or the limits of the device. Thanks to everyone who built with us, and to the Qwen and Alibaba Cloud teams for hosting. All nine project repositories are linked above. Building on SIE? [Star the repo](https://github.com/superlinked/sie). ![Teams presenting demos at the end of the hackathon day](/images/blog/qwen-hackathon/hackathon-demos-room.jpg)