Qwen3.5-0.8B quiet assistant v1 (q4f16_1, MLC / WebLLM)
A small in-browser assistant for complex work apps, without a chat box. It runs in the background with WebLLM: after each pause in the user's work it calls the app's tools one after another (ReAct), and the app shows what the calls produce (the next step, a filled form, an insight, a summary, the day's overview). The model never writes text: the app's code writes every word and number from its data. Download: 0.43 GB.
This is the f16 build: it needs the WebGPU shader-f16 feature. The f32 build has the same weights and runs on any WebGPU. Other build: tfukkwang/Qwen3.5-0.8B-quiet-assistant-v1-q4f32_1-MLC.
What it does
Its context always holds the user (new or experienced), the time, the place, their habits, the recent events and the screen. It picks from the same tool names in every app: queue, day_overview, open_item, item, playbook, guide_next_step, show_insight, summarize, autofill, open_tab, done, plus each app's own lookups. Force one tool call per step with a grammar (WebLLM structural tags).
Training
- Start: the Qwen3.5-0.8B tool router v3 (tfukkwang/Qwen3.5-0.8B-tool-router-v3-q4f32_1-MLC). No SFT.
- Method: GRPO with the DAPO loss (token-level loss, clip-higher 0.28, no KL), LoRA r=16, 150 steps, 8 samples per prompt, on 895 of 1,114 simulated moments whose 8 rewards differed. One RTX 3090, about 1 h.
- Data: moments simulated in three apps (claims desk, purchase-to-pay, underwriting workbench) with the apps' real code: the prompt is exactly what the browser sends; the reward of every possible next call comes from the app code (help that fits the person, what they asked first, insight topics with real findings, at most two cards, then done; errors and repeats cost).
Results (greedy, no grammar; share of steps where the model picked the best call)
|
Start (router v3) |
This model |
| Clinic front office (an app never in training), 389 steps |
31% |
67% |
| New moments in the training apps, 134 steps |
31% |
60% |
| Invalid calls, clinic |
13% |
2% |
Known weakness: for an experienced user on an open item it still often shows the next step instead of an insight (48% best on clinic).
