Announcing EVMDec-3: Our Latest Smart Contract Decompiler Model
We've been heads-down on EVMDec-3 for months, pairing bench testing with real decompiles from folks in the community. Today we finally get to share it: PLUS and PRO subscribers can start using the new model right away, and it raises the bar for how readable AI-assisted decompilation can feel.
This is our first release with full benchmark notes attached. Every number below comes from EVMDec-Bench v1 — a locked dataset of 46,040 Ethereum contracts the model never saw while training — so you can follow our progress run to run.
Evaluation Metrics
We kept the scorecard simple. Four things matter most when you ask us to turn bytecode back into something you can reason about:
Completeness
Does the model actually finish the job? We track whether every contract comes back with a full reconstruction instead of a clipped stub.
Partial snippets can still spark ideas, but consistency here tells us the model is sturdy.
Recompilation
We push the generated Solidity back through a compiler. If it turns into valid bytecode, we know the structure wasn't just pretty — it was right.
ABI Recovery (Functions)
When recompilation works, we check how many function selectors and signatures line up with the original contract.
Functions are where most of the day-to-day interaction happens, so we track them first before expanding to events and errors.
Textual Similarity
Finally, we look at how the language itself compares. A quick TF-IDF pass tells us whether the rewritten contract sounds like something a human actually published onchain.
Results
Completeness
- 100% completeness: every one of the 46,040 contracts finished rendering.
- No truncated or malformed results showed up in this run.
Recompilation
🎯 Industry first for real: EVMDecompiler is still the only decompiler we know of that lets you round-trip back into valid bytecode.
- 49.8% success rate: almost half of the contracts slot straight back into the compiler and pass.
- It's a strict test — tiny typing mismatches are enough to fail — so flirting with 50% is a solid base we plan to build on.
- And yes, we keep checking: nobody else offers recompilation today.
ABI Recovery
Once we have a compiling contract, ABI recovery tells us how faithful the functions look.
- Average ABI match rate: 31.1%.
- Median ABI match rate: 13.0%.
- 19.1% of contracts clear the 95% mark, and a healthy slice land on perfect matches.
ABI Recovery Distribution (CDF)
Textual Similarity
On the language front, most reconstructions read like the original authors:
- Average similarity: 0.834.
- Median similarity: 0.939.
- 46.2% of contracts score north of 0.95.
Textual Similarity Distribution (CDF)
Availability
Ready to try it? Here's how access looks today:
- PLUS users: EVMDec-3 Model (256K tokens)
- PRO users: EVMDec-3 Model (256K tokens)
Key Takeaways
EVMDec-3 clears the benchmark with 100% completeness.
Recompilation lands at nearly 50%, showing the structure isn't just cosmetic.
ABI recovery is strong on a meaningful slice of contracts, with plenty of perfect matches.
Textual similarity stays high, so the code feels readable without heavy cleanup.
With EVMDec-3 and EVMDec-Bench v1 in place, we finally have a baseline we trust. Every future release will push against the same dataset so you can see progress without reading between the lines.