The July-Aug number IMO is too high. The recent market-moving news is just the release of V3.1 with some reasoning improvements and a larger context window. There are only 13 days left in August.

@ookina_inu

>"Does not have to be named V4 explicitly; would be whatever model that succeeds V3, probably whatever they end up naming the model behind the deepseek-chat endpoint"

@notarealuser I would be extremely surprised if this were resolved yes on the basis of the V3.1 drop alone (just as GPT 4.1, 4.5, etc did not satisfy GPT-5 markets). Categories are fuzzy, but to me the V3.1 naming clearly implies that it’s an incremental update to V3, not a different model. I imagine at the very least the architecture, parameter count, etc would have to be different for the new model category to trigger.

@ManifoldAI

@ookina_inu I placed a bet based on resolution criteria. Obviously it's an incremental model, bigger context

@notarealuser

1. A longer context length does not imply a different model. Context length is one of the easier things to change about an LLM. Labs often serve different context length versions of the exact same model.

2. This is IMO too literal a reading of the market description. The spirit of the question is about the next major version of model release by Deepseek. By your logic, the V3-0324 and R1-0528 updates would have counted as new models if they had been released while this market was open. The wording in the description was clearly meant to catch the scenario where DS releases a V4-equivalent (new architecture, etc) but calls it something different, like “AGI-1” or something.

3. It is still overwhelmingly likely that Deepseek will release a new model literally called “V4” in the near future. It would be pretty silly to resolve a question about V4 to “August” on the basis of V3.1, only to have the actual V4 come out in (without loss of generality) September.

4. Again, it’s the exact same 671B architecture as all the previous variants of V3. No NSA, no parameter count change, etc. It’s a new checkpoint on top of the same base. There is a broad understanding that labs can and do ship new checkpoints of the same model, and that this is a qualitatively different phenomenon than shipping a new model.

deleted

arb