DeepSeek has begun quietly testing V4.1-Flash inside its web chat, swapping in a variant tuned for tool-calling agents that spits out tidy lists and structured steps instead of chatty paragraphs, especially on intricate coding prompts like raymarching demos.
Yeah it seems we have DeepSeek-V4.1-Flash on web now. Announcement within hours-days I'd guess very different style (old vs new). And I think it's no good for web – when I gave it a console error, it Counted Every Line "manually". It is designed as a tool-calling agent first.
@teortaxesTex Well I wonder if V4-Instant is also dumber now.
Side-by-side tests show the model favoring concise, numbered outputs that slot straight into agent workflows, trimming tokens while keeping the focus on executable steps rather than explanatory fluff.
No confirmation yet on whether V4.1-Flash will reach the public API or expand beyond the current gray-scale web test, leaving timing and access details open.
Positive users praise DeepSeek V4.1-Flash for greater token efficiency and as good news with future potential, while negative users call the console error handling brutal and poorly optimized for conversation.
@teortaxesTex Looks like they are focusing on token efficiency now
Yeah it seems we have DeepSeek-V4.1-Flash on web now. Announcement within hours-days I'd guess very different style (old vs new). And I think it's no good for web – when I gave it a console error, it Counted Every Line "manually". It is designed as a tool-calling agent first.
@teortaxesTex So it is coming 🤔 this is good news in general hopefully the vision also will be available.
@teortaxesTex Kinda weird, I gave it some random three js prompt and instead it generate a text description of the prompt lol
@teortaxesTex maybe announcement is still 2 months away like last time🤣
@teortaxesTex Much more token efficient and resembles GLM
@teortaxesTex Even the Expert thinking is different now
@teortaxesTex So excited
@teortaxesTex looks just like glm models cot
@teortaxesTex so no 4.1 pro update? 😅
@teortaxesTex so a regression?
@teortaxesTex Counting lines manually for a console error is brutal
feels like it optimized for everything except conversation