Train like a flock
FLock은 왜 이름이 FLock(새 떼)일까요?
병원, 정부, 기업은 저마다 좋은 데이터를 갖고 있지만 밖으로 꺼낼 수 없어요.
그래서 @flock_io는 데이터를 옮기지 않고 모델이 찾아가는 연합학습을 써요.
1. 각 기관이 자기 데이터로 직접 학습
2. 데이터는 그대로 두고, 배운 결과만 모음
3. 모아서 하나의 모델로 합치고 다시 돌려보냄
이걸 반복할수록 모두의 모델이 점점 좋아져요.
실제로 하고 있는 일
- Red Hat과 협업: 정부·기관에 연합학습 제공
- UNDP와 MoU: 도미니카공화국 공공 부문 디지털 혁신
- THIS/THAT 모델 공개: 1.88B 파라미터, 의사결정 과제 1,710개에서 성공률 87.8%
데이터는 집에 두고, 배운 것만 날아올라요.
FLock.io retweeted
FLock은 왜 이름이 FLock(새 떼)일까요?
병원, 정부, 기업은 저마다 좋은 데이터를 갖고 있지만 밖으로 꺼낼 수 없어요.
그래서 @flock_io는 데이터를 옮기지 않고 모델이 찾아가는 연합학습을 써요.
1. 각 기관이 자기 데이터로 직접 학습
2. 데이터는 그대로 두고, 배운 결과만 모음
3. 모아서 하나의 모델로 합치고 다시 돌려보냄
이걸 반복할수록 모두의 모델이 점점 좋아져요.
실제로 하고 있는 일
- Red Hat과 협업: 정부·기관에 연합학습 제공
- UNDP와 MoU: 도미니카공화국 공공 부문 디지털 혁신
- THIS/THAT 모델 공개: 1.88B 파라미터, 의사결정 과제 1,710개에서 성공률 87.8%
데이터는 집에 두고, 배운 것만 날아올라요.
FLock.io retweeted
Every time we use a mainstream AI tool, we help train it, and pay to do so. 🤖
Tiffany Wang of @flock_io works on decentralized AI, where the people who help train a model own a piece of it, and communities keep control of their own data.
cryptoaltruists.com/blog/cry…
The marketplace is the easy half.
The hard half is supply. A specialized model is only as good as the proprietary data behind it, and the most valuable data (clinical, financial, legal, genomic, industrial logs, and more) can't legally be pooled.
Federated learning changes that. Models train where the data lives, and only updates move.
That turns data silos into collaborators: labs, companies, institutions, hospitals, banks, or ministries can co-train a model none could build alone, without handing over raw data.
Open-weight specialists also need an open coordination layer for training.
That's what FLock can do.
To decentralize AI and preserve freedom, we need a coordination layer: a market economy where agents communicate and cooperate through prices.
Teams of specialized open-weight models can outperform frontier models, just as a team of skilled people can outperform any single person. When those teams form companies, specialize, and coordinate across an entire economy, you get civilizational progress.
Coinbase is building the financial rails and accounts for the agentic economy. We’re also building the marketplace where agents can find, hire, and pay one another (see follow on post below).
Every company can now sell its own specialized open-weight model, trained on its proprietary data, that beats the frontier on narrow, high-value tasks. Builders can shop the marketplace and assemble teams of these specialized agents in minutes.
Freedom of choice and real competition in the agentic market will create far better outcomes for everyone.
Dubai in December.
FLock at @GITEX_GLOBAL, UK Pavilion.
Private AI for businesses whose data can't leave the building.
Come say hi. Bring your hardest yes/no question.
FLock it.
nitter.cf/flock_io/status/212231…
Bringing private AI to real-world businesses.
We’ll be at @GITEX_GLOBAL in Dubai this December as part of the UK Pavilion, organised by @techUK, @biztradegovuk, and Tradefair, showcasing our privacy-first enterprise AI solutions.
Who else will be there?
Spending on specialised AI models is expected to double this year. Why?
- Frontier AI models come with rising costs
- Businesses need speed and reliability for specific tasks
- Sensitive data needs to stay private and under their control
The AI model you own could become your next competitive edge.
Train your own with FLock.
Bringing private AI to real-world businesses.
We’ll be at @GITEX_GLOBAL in Dubai this December as part of the UK Pavilion, organised by @techUK, @biztradegovuk, and Tradefair, showcasing our privacy-first enterprise AI solutions.
Who else will be there?
A 9B model, fine-tuned for under $100, beat Opus 4.5 at fraud detection.
Fraud check isn't an essay. It's a decision.
THIS / THAT: 1.88B. Runs locally.
One typed answer + a confidence score.
Still paying frontier prices to say yes or no?
FLock it.
nitter.cf/Acurast/status/2118719…
A free 9B model, fine-tuned by @coinbase for under $100, beat Opus 4.5 at fraud detection.
Acurast runs that same model (Qwen3.5-9B) on a phone.
How fast is it? What does it cost? How does it scale to 100+ phones?
Read the full test results and deploy it yourself 👇
acurast.com/blog/network-and…
Enterprise private AI is only as strong as the research underneath it.
FLock's Chief AI Scientist @zehuac presented three FLock papers as oral presentations at CIKM 2026 in Rome. Each one tackles a challenge in making private AI practical for production at scale.
- Can we train models on our own sensitive data?
- Can we afford the training?
- Can we trust what the model tells us?
1. Anonymisation that keeps data usable
Graph-Driven Context-Preserving Anonymization removes every identity and each reference to it while keeping the reasoning intact. On clinical data, it removed 100% of annotated identifiers, with accuracy within 0.1 points of the original.
2. Cheaper Transformer training
Gradient Interaction Graph Optimization (GIGO) learns which parts of a model depend on each other and updates them in that order. It reaches the same quality in 32 to 34% fewer training steps than AdamW.
3. Tighter uncertainty guarantees
Cavity-Calibrated Prediction Sets (CCPS) give structured outputs a confidence guarantee an auditor can check, with answer sets 34 to 58% smaller than the standard approach.
More enterprises are building AI that runs inside their own walls as a strategic advantage. The hard parts are using sensitive data safely, keeping compute costs down, and producing verifiable outputs. Our research is focused on solving them.
Private AI is moving from the lab into production. We're building for that shift.
Huge congratulations to all the authors: our Chief AI Scientist @zehuac, Wei Dai, our CEO and founder @0x7SUN and Jianwei Yang.
Read the full papers ↓
Gradient Interaction Graph Optimization
→ doi.org/10.1145/3799682.3840…
1
2
199
Cavity-Calibrated Prediction Sets dl.acm.org/doi/10.1145/37996…
1
151
FLock.io retweeted
API 키 하나로 모든 모델을 🔑
FLock API 플랫폼 하나면
✅ DeepSeek · Gemini · GLM · Kimi · Qwen 등 16개 모델
✅ OpenAI 코드 그대로, base URL만 Flock으로
✅ 공급자가 느려지면 자동 우회 (스마트 라우팅)
✅ 로그 · 사용량 · 팀 관리 · Playground
✅ 크레딧 충전식 (Stripe · Base Pay · Coinbase)
FLock의 THIS/THAT 결정 모델은 $0 ⚡
👉 platform.flock.io
Why juggle multiple AI providers when you can access them all in one place?
One key for top open-source models from @deepseek_ai, @Alibaba_Qwen, @Kimi_Moonshot, @Zai_org, @GeminiApp, and FLock.
One platform. Endless possibilities.
Build with FLock.






































