We just doubled our speculative code edit throughput and hit 10k tok/sec per request!
Morph now merges code at 10,500 tokens/sec — roughly 4× faster than the best speeds on Cerebras.
That kind of speed makes previously impractical workloads trivial: applying complex edits across a 40k-token document now takes under 4 seconds. This isn’t a vanity metric - we think it unlocks an entirely new domain of AI use cases where codebases, configs, or long documents can be semantically edited in real time.
Morph is a Fast Apply model dedicated to merging edits from frontier LLMs
We want to enable developers to build realtime interfaces with AI
Help me understand. Is this for cases where you have a file and you "ask" an LLM to change something, and they reply in chat mode with something like < //--unchanged code \n changed line \n changed line \n //----remaining code unchanged > ?
If so, isn't this flow like 6mo old, and not really used anymore? The latest tools (terminal based and vscode extensions like cline/roo/kilo) already support "diff edits", where the model outputs a diff format that the tool speaks. I get "instant" edits that way, right in my IDE, and model support has been great (gpt5,claude4,gemini2.5,grok-fast-1, etc.)
So what's the use case of this model, then? Cool technical results, and congrats, but it seems the "field" has already solved for this particular problem?
Using fast apply is more reliable at first pass and is faster/cheaper. You prompt your agent to output in a lazy format and our model learns hoe to merge it in.
The ides listed typically do turn based search and replace which uses more tokens and takes longer
This weirdly seems like its the best mechanism to buy this much data.
Imagine going to 500k publishers to buy it individually. 3k per book is way cheaper. The copyright system is turning into a data marketplace in front of our eyes
I suspect you could acquire and scan every readily purchasable book for much less than $3k each. Scanhouse for instance charges $0.15 per page for regular unbound (disassembled) books, plus $0.25 for supervised OCR, plus another dollar if the formatting is especially complex; this comes out to maybe $200-300 for a typical book. Acquiring, shipping, and disposing of them all would of course cost more, but not thousands more.
The main cost of doing this would be the time - even if you bought up all the available scanning capacity it would probably take months. In the meantime your competition who just torrented everything would have more high-quality training data than you. There are probably also a fair number of books in libgen which are out of print and difficult to find used.
psychologically bad for a person isn't grounds for something arising. I'm not arguing this should or shouldnt be morally, I'm just observing the situation.
Utility to society doesnt always account for the survival of society
it really matters when you care deeply about the product experience. The gap between something that works sometimes and something that works great everytime is what separated Cursor from the rest.
