A new reasoning model release doesn't grab my attention as much as the fact that it's a Mamba-based model that's competitive with conventional architectures.
(Strictly speaking I'm a bit late, because it seems the non-reasoning version was released earlier in March, but I didn't see it until today).
I just gave it a unit test I was in the middle writing and asked it to fix it - no context/documentation/interfaces, literally just the test. It performed amazingly well - even where it didn't get it right, it made a very reasonable guess (e.g. calling dispose() when the actual method was destroy() - but since I hadn't given the interface, it had no way of knowing).
I'm going to road-test this a bit further via Cline (where I alternate between Gemini Flash and DeepSeek).
I'm surprised this isn't getting more attention here. This may well be the future.
These large hybrid transformer-SSM models are faster and more efficient but it seems not really by a huge amount: a big deal but Moore's law progression level not breakthrough level.
I gave it a tricky maths problem (from AIMO) and it spat out about 25k thinking tokens (my estimate), going in circles, which took many minutes, far longer than a better model, until finally figuring out the right answer. Not sure whether that's evidence that it is or isn't weaker when looking beyond recent context.
3 comments
[ 3.2 ms ] story [ 10.6 ms ] thread(Strictly speaking I'm a bit late, because it seems the non-reasoning version was released earlier in March, but I didn't see it until today).
I just gave it a unit test I was in the middle writing and asked it to fix it - no context/documentation/interfaces, literally just the test. It performed amazingly well - even where it didn't get it right, it made a very reasonable guess (e.g. calling dispose() when the actual method was destroy() - but since I hadn't given the interface, it had no way of knowing).
I'm going to road-test this a bit further via Cline (where I alternate between Gemini Flash and DeepSeek).
These large hybrid transformer-SSM models are faster and more efficient but it seems not really by a huge amount: a big deal but Moore's law progression level not breakthrough level.
I gave it a tricky maths problem (from AIMO) and it spat out about 25k thinking tokens (my estimate), going in circles, which took many minutes, far longer than a better model, until finally figuring out the right answer. Not sure whether that's evidence that it is or isn't weaker when looking beyond recent context.