How WorkBuddy Hit 20M MAU in Months: Two All-Nighters, Engineering Over Model Strength

WorkBuddy, a Tencent AI workspace product, emerged not from a top-down strategic mandate but from a ten-person AI coding assistant team that spent weeks on-site at a bank learning security sandboxes and permission boundaries. That groundwork became the foundation for a product that would later reach 20 million monthly active users.

① Origin: A Weekend, Two All-Nighters, Version 0.01

In mid-January 2026, team lead Wang Shengjie and an operations colleague built the 0.01 version in two all-nighters: a minimal chat interface preloaded with curated Skills, ready to use out of the box. No formal project approval — just a demo video sent to leadership, approved the same night, internal beta launched Monday. By March, over 2,000 non-technical Tencent employees (HR, admin, operations, sales) were daily users, proving the platform could be adopted by people who don't write code.

② Climb: Three Months, 40+ Versions, Team 10→100+

Public launch on March 9, 2026 triggered demand that overwhelmed capacity — requests exceeded CodeBuddy by multiples, forcing a 10x emergency scale-up within hours. The team iterated nearly daily (40+ versions in three months, roughly one every two days) and grew from 10 to 100+ people. Monthly visits rose from 8.85M in March to 20.97M in June (2.4x growth), capturing one-third of the entire desktop AI office agent market (60M total per Analysys). Unlike Codex or Claude Code which require environment setup, WorkBuddy eliminated local deployment, network, account, and permission complexity — download, open, work. This low floor enabled rapid penetration into enterprise IT without extra security audits.

③ Pivot: Hunyuan Hy3 Integration — Success Rate 72%→90%, Latency -34%

On July 6, 2026, WorkBuddy integrated Tencent's Hunyuan Hy3 (MoE, 295B total params, 21B active, 256K context). Internal evals showed office task solve rate jumping from 72% to 90% (1-in-4 failures to 1-in-10), average completion time down 34%, and user self-selection of Hy3 up 6x. Hy3's "slow-thinking" approach — understanding the end goal before decomposing steps — replaced the old "outline-only" behavior. The two-week free tier served a dual purpose: user acquisition and, more critically, collecting first-hand success/failure trajectories to feed reinforcement learning.

④ Retention Evidence: DAU/MAU 65%-75% vs Industry ~20%

Analysys data shows WorkBuddy's DAU/MAU ratio sustained at 65-75%, comparable to Slack, while domestic AI apps average ~20%. High ratio means users treat it as a daily workbench, not an occasional toy. Citi Research called this "remarkably strong" for a months-old product. The reason is pragmatic: target users (HR, admin, ops, sales) care only "did the task finish?" Once they deliver weekly reports, spreadsheets, and analyses through it consecutively, it becomes a workflow fixture.

⑤ Consumption Curve: Per-User Daily Tokens 10x in 3 Months, Then 3x More

Liu Yi (CodeBuddy & WorkBuddy lead) disclosed at Tencent Cloud AI Industry Conference: per-user token depth grew from ~50-60K in March to ~550K in June (10x), then ~1.67M in August (3x again). Task complexity escalated from "rewrite a sentence" to "cross-table reconciliation + WeCom distribution + multi-agent parallelism." Combined WB+CB daily token burn rose from 0.1-0.15 trillion (March) to 0.9 trillion (June) to 6.5 trillion (August median). By early August, WorkBuddy MAU ~14.7M, DAU ~3.9M, burning 5-8 trillion tokens daily. This reflects a positive feedback loop: each successful task emboldens users to delegate larger ones.

⑥ Unsolved: Compute Saturation, >50% Queue Rate, Still a Cost Center

The July 8 "free-tier incident" exposed infrastructure limits: Hy3 launch spiked consumption to peak by 10 AM, queue rate exceeded 50% by afternoon (one in two users waiting for compute). Joint team scrambled overnight; capacity restored next day, free tier extended to Aug 5. For an efficiency tool, wait time directly erodes core UX. Deeper issue: elastic infrastructure insufficient for demand spikes. On profitability, Pony Ma's deputy James Mitchell (汤道生) stated WorkBuddy has no commercialization KPI yet — mature businesses (up 17% YoY ex-AI) fund the burn. C-end paid rate ~3.2%; real revenue is B-end (¥198/user/month, 140k+ paid seats), but enterprise renewal rates remain unproven.

WorkBuddy teaches the industry: AI application breakout may come not from "strongest model" but from "right engineering, smooth ecosystem, light organization." The 0.01 bet: ordinary people need AI that works, not just chats. The 40-version bet: "download and run" beats "higher capability ceiling." The Hy3 bet: product-data flywheel feeding model improvement. Today's compute and cost crisis is the next hurdle. From 0.01 to 20M MAU proves "it can run." Turning cost center into profit engine is the real second half.