DeepSeek V4.1 Flash

“Used it in dsh, it works and I review, then submit requirements and feedback and directly insert into the reply, but the result is dsh keeps fixing the latest feedback, forgetting all the previous requirements.”

“The API also started refusing, but when I clicked regenerate output, it wrote it out.”

“It basically doesn't even self-test on technology stacks it's not very familiar with, only doing simple static compilation checks and calling it done.”

“4.1-flash is fast and good at fixing problems with clear-cut correct answers (like UI errors), but for problems with complex requirements and fuzzy results (like permissions), it will bury tons of pitfalls, actually regressing compared to the previous version of flash”

“Anyway, 4.1 really found bugs that astra couldn't find, and the speed was ridiculously fast”

“The TPS looks high, but the time to solve individual problems is about the same as 4 flash, feels like the thinking chain got longer”

“DeepSeek v4.1 Flash is already on par with Fable 5. It feels like a lifetime has passed. The once earth-shattering FatBoy is now just another nobody, sharing a table with cost-effective models like Flash.”

“The Emperor's Uncle can still output without pressure, any style works, just a bit formulaic”

“Deepseek flash was insanely fast for me when I used it.”

“Today DeepSeek V4.1 flash was first at solving a rather simple UI bug in a vibe coded game of mine”

“Did a private benchmark suite and saw that Opus 5 medium and Deepseek V4.1 Flash high were on par as reliable coding agents, with the small catch that Opus cost 30x more.”

“Just a pure test-scoring machine, on the same level as Muse. When it comes to actually using it for real work, it's worthless”

“From my personal experience, DeepSeek-V4.1-Flash is worse than GLM-5.3-Flash”

“After adding the engram plugin, there's a devious workaround—you can directly update the knowledge base.”

“I discussed their PDF with K3 for a few rounds, and the conclusion is that there are still some tradeoffs—particularly, long-text understanding may have structural issues to watch out for. In other aspects, world knowledge and deep domain knowledge are just a bit behind, but for a Flash model, it's already quite competitive.”

“Now the armor is too thick”

“At least from my testing, the writing style in RP scenarios isn't as good as v4p's.”

“I used v4.1 flash both with Claude Code plugin connected to DeepSeek API and in CodeBuddy, they feel about the same, both very satisfying”

“The price is very flash, but the performance is not flash at all.”

“it dominates the Pareto frontier, only 4 model here above it, rest below.... Fucking killed an entire range of models”