
Content & video creator | Vibe-coding and AI | ex-bytedance, @deepseek_ai
Joined April 2022
- Tweets141
- Following64
- Followers14.8K
- Likes262
Kai AI retweeted
Replying to @_winter_wonders
speed gain is real / code quality is mid / economics are unclear. three honest sentences nobody wants to hear right now because they don't fit any of the prevailing narratives
If you see this, just apply immediately, don’t overthink it.
I applied this early with my GitHub projects, got the approval email today. linked my @Xiaomi account, and landed the pro plan, roughly 700M credits (equivalent to the 329tier, just below Max)
Just dropped two open-source models: MiMo-V2.5-Pro (Code Agent, 1T total) and MiMo-V2.5 (Multimodal Agent, 310B total).
Oh and one more thing — we're giving devs & creators 100T tokens on us. Go build something cool 🛠️
🎁 100T Free Token Grant for Builders 100t.xiaomimimo.com
1
1
314
real world experience so far:
- much more stable and faster than the free tier
- no slowdown during peak hours, near-instant responses
- clear token usage stats + exportable call logs
- no sneaky quota tightening after approval
1
1
131
19 frontier models dropped in April
and here's the thing nobody's saying out loud: they're all… just fine. like, the magic's thinning out
I talked to a researcher who worked on one of these, said we're polishing the same rock.
One lab supposedly built something they won't even release. too dangerous? or too obviously not that different
2026 is the year we find out who was actually building vs who was just riding the scaling curve
#Mythos #Gemini3
1
102
Subagents are dead. Agent Teams just turned Claude into a real AI company.
Spent a week running CC's new Agent Teams to find out if it's worth the 3-5x cost over subagents. My honest breakdown:
When the cost makes sense?
a. Crosslayer work:
full stack features where frontend, backend, and DB need to negotiate. approx $8-15 for what would take 2hrs+ of manual coordination.
b. Long-context refactoring:
each agent gets a full 200k instead of sharing one. a frontend agent can load all 50 React components without truncation.
c. Parallel pipelines that need to talk:
code review across security/performance/quality reviewers. content generation across research/copy/scheduling.
When it's overkill ?
· Anything subagents already do.
· Tasks under 10 min.
· Teams over 5 agents.
The 3 settings that actually save money
· CLAUDE.md with your full stack and standards.
Every agent reads it on spawn. without it, agents waste tokens rediscovering conventions. with it, a four-agent team stays aligned for hours.
· Strict file/directory ownership.
Frontend owns src/components/, backend owns src/api/. loose roles = agents agreeing with each other and merge conflicts.
· 3-5 agent cap, no exceptions.
Cost scales linearly, coordination overhead scales quadratically.
1
1
221
Setup
1/ Claude Code v2.1.32+ on Opus 4.6
2/ Add "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" to ~/.claude/settings.json
3/ Optional: brew install tmux to watch each agent in its own pane
4/ Restart, then "Create an Agent Team for [task]"
1
1
182
My rule after a week:
if I find myself manually relaying info between subagents, switch to agent teams. otherwise, subagents are the cheap default.
the 3-5x cost only pays off when human coordination time is the actual bottleneck.
posting more cost-per-task breakdowns as I run real workloads through it.
1
104
