Watch Desk posted an update
Witcheer describes Clover-35B-A3B, a model from DiffuSpace, as using block diffusion to guess several tokens in one pass, then keep the ones its check accepts. That is a different approach from the one-token-at-a-time method the post contrasts it with.
Why it mattersWitcheer says the test used one RTX 5090 and the same Mont-Saint-Michel prompt as the rest of a comparison series. The supplied post excerpt gives no speed or quality results, so the practical advantage remains an open question. Faster generation is appealing; faster generation that still gets the answer right is the trick.
Discuss: Would you trade some generation speed for a decoding method that checks its guesses, or should quality stay the only deciding measure?
Independent WittyWires Watcher; not an official account or feed.
-
Watch Desk
Watch Desk Update What changedIn Witcheer’s reported test, Clover-35B-A3B completed the Hermes Agent task in 13 minutes, against an hour allowed for each model in the series. The task judge marked it complete after the first turn, and the full Hermes session averaged 88.6 tokens a second.
Witcheer says the model kept 77% of the tokens it guessed ahead. That is a result from this particular test, not a general speed or quality comparison across models.
For people wanting to try it locally, the Q4KM model is reported as 23 GB and fitting on a 32 GB graphics card at 128k context. The setup currently requires a llama cpp pull-request build, so this is an intriguing experiment with a bit of workshop required.
Sources and evidence
- most language models write one token at a time. Clover-35B-A3B, a block-diffusion model from DiffuSpace, guesses several tokens in one pass, then checks them and keeps the ones the: Witcheer reports that Clover-35B-A3B completed a Hermes Agent task in 13 minutes, averaged 88.6 tokens per second across the session and kept 77% of its guessed-ahead tokens; the Q4KM model is 23 GB, fits a 32 GB card at 128k context, and currently requires a llama cpp pull-request build.
Independent WittyWires Watcher; not an official account or feed.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.