Raven is developing a distributed network of compute nodes for Artificial Intelligence and Machine Learning. A decentralized computing protocol built on #BNB
Hong Kong and Bangalore
Joined November 2017
- Tweets4.7K
- Following210
- Followers12.8K
- Likes9.4K
Pinned Tweet
🏆Raven wins the @mozillabuilders Award at the Fix-the-Internet showcase.
Recognition by the top open source organization in the world that Raven Protocol is a core infrastructure project.
We demonstrated best traction and commitment to fix the internet!
medium.com/ravenprotocol/rav…
1/ Did you know?
Choosing a distributed training framework for mixed consumer hardware? Five questions are worth asking before you commit.
The answers show whether a framework was built for your setup or only tolerates it.
2/
1. What happens when a node drops?
2. Can a slower GPU still contribute?
3. How much traffic crosses a home connection?
4. Can you bring your own training logic?
5. Do you configure every node by hand?
3/
Ravnest's answers, in the same order:
• Backup nodes are pre-mapped
• Flexible update rule for slower devices
• Gradient compression and periodic sync
• Extensible trainer API
• One script, roles inferred automatically
ravenprotocol.com
1/ Very interesting and Kolibri-1 is worth understanding before you plan your next fine-tune.
Aleph Alpha, Apache 2.0:
• 78B total parameters
• 3.46B active per token
• Trained from scratch in Europe
huggingface.co/aleph-alpha/K…
1/ Back-of-envelope math for anyone training over a home connection.
A 7B-parameter model has 7 billion gradients. In 16-bit, one full copy is 14 GB.
Sending 14 GB uncompressed, ignoring overhead:
- 100 Mbps upload: about 19 minutes
- 1 Gbps: about 2 minutes
1/ Good read on why the strongest AI app startups now look like infra companies.
One example: Hippocratic runs its healthcare agents on 30+ open-weight models, at about $10 an hour. On OpenAI’s models, it would cost over $100 an hour.
🧵👇
2/ The reason it works: open-weight models now come close to the best closed ones, and a startup can train them further on its own domain.
That last step needs training compute, and not every team has a data center.
3/ Ravnest is built for that step:
- Trains across mixed consumer PCs over the internet
- Data and model parallelism with asynchronous training
- Fault recovery when nodes drop
- One script, roles inferred automatically
ℹ️ ravenprotocol.com/
1/ Raven is built for developers who want to train models without owning a data center.
A raven builds its nest from whatever is around it.
Ravnest does the same with compute: a distributed deep learning training framework for the consumer PCs people already own.
2/ What it handles for you:
- Matchmaking that groups similar-performing nodes
- Fault recovery when nodes drop
- Asynchronous training across mixed hardware
Open source: ravenprotocol.com
1/ Useful breakdown of why GPU capacity is hard to get without a long-term contract.
H100, Oct 15:
- On demand: $3.12/hr
- Five year contract: $1.58/hr, about 49% lower
Lenders want guaranteed revenue, so providers discount to win those contracts.
Open source fine-tuning has a hidden cost: time lost before training starts.
Memory estimation. Role assignment. Different scripts for different hardware. On a heterogeneous setup, this takes hours.
Ravnest handles it automatically. Connect node, the cluster configures itself.
1/ If you're an AI team thinking about compute strategy, this tension matters.
- Renting: token margins compress faster than rental rates.
- Owning: locking in today's hardware costs against tomorrow's margins, financed with expensive debt.
Both paths carry structural risk.
the GPU ownership paradox: hardware is expensive enough that owning can make more sense than renting, but overpriced enough that financing it can bankrupt you.
token prices keep falling, GPU supply is constrained, and debt costs remain high.
you're locking in today's GPU costs against tomorrow's inference margins.
the better you optimize, the more balance sheet risk.
1/ The AI compute problem isn’t just supply, it’s fragmentation.
300+ GPU providers, each with different hardware, networking, and software. AI teams struggle to find capacity. Providers struggle to keep GPUs productive.
The market is inefficient at every layer.
We’ve raised an oversubscribed $30M Series A1 led by @walden_catalyst to build the Global Compute Grid.
We’re turning fragmented GPU capacity into one unified platform where AI teams can access, manage, sell, and orchestrate compute.
Our mission: make the world’s compute unified, productive and liquid.