@YRSM_Simon: Don't be fooled by the name "Max". I tested the same task 100 times. deepseek v4 flash max is 26% faster than medium, saves 34% tokens, and scores 83% higher on quality. Although max thinks more, it also requires fewer rounds. (Every task is different...
Summary
The author tested the same task 100 times and found that DeepSeek V4 Flash Max is 26% faster than the Medium version, saves 34% tokens, and scores 83% higher on quality. Although Max thinks more, it requires fewer rounds.
View Cached Full Text
Cached at: 08/14/26, 09:36 AM
Don’t be fooled by the name “Max”.
I tested the same task 100 times.
deepseek v4 flash max is 26% faster than medium, uses 34% fewer tokens, and scores 83% higher on quality.
Although max thinks more, it also needs fewer rounds.
(Every task is different; you’ll need to test the specific configuration yourself) https://t.co/ALj5llIOf6
Similar Articles
@jakevin7: DeepSeek cache hit rate 95%, feels great. Maka's performance under the latest round of long-context tasks with the Deepseek model is outstanding. Total runtime close to 18 hours, nearly 400 million tokens, cost 33 bucks. The Make builders are amazing…
DeepSeek cache hit up to 95%, Maka desktop AI workstation performs excellently in long-context tasks, supports multiple models and tools, open source and local-first.
@FuckAnthropic: Conducted a comparative analysis. Overall, DeepSeek V4 Flash-0731 is roughly a model at the level between Opus 4.7 and 4.8, entering the frontier Agent model competition with a minimal activation scale, and at about 1/12 to 1/60 of the token cost to enter the frontier Ag…
The author's comparative analysis concludes that DeepSeek V4 Flash-0731 achieves Opus 4.7–4.8 level performance with an extremely small activation scale, entering the frontier agent model tier at a very low token cost. It surpasses GLM-5.2 overall, but its shortfalls remain difficult repository-level coding and long-horizon engineering.
@jakevin7: Maka's Harness project brings DeepSeek Flash's test set results close to GLM-5.2 level ----------------------------------- maka + DeepSeek Flash V4, te…
Maka's Harness project improved the self-check mechanism, enabling DeepSeek Flash V4 to achieve evaluation results close to GLM-5.2 on the terminal-bench sample set, completing 10 programming agent tasks with only 4 RMB and a 97.5% cache hit rate.
@FinanceYF5: 1/ Someone ran 3 tasks with Fable 5. The $200 Max membership, 5-hour quota depleted 73%. He said he was stunned. Not because it's expensive — because Opus 4.8 never produced anything like this. A thread explains clearly
A user ran 3 tasks using the Fable 5 model, consuming 73% of the Max membership's 5-hour quota, sparking discussions as Opus 4.8 never consumed resources in this way.
@YRSM_Simon: Let Claude Fable 5 play chess against locally running DeepSeek V4 Flash (DGX Spark, thinking maxed out) — the year's most absurd comparison: Fable: ~2,600 tokens per move, 50 seconds to move; V4 Flas…
The author had Claude Fable 5 play chess against a locally running DeepSeek V4 Flash. As a result, DeepSeek consumed tens of thousands of tokens per move and even timed out under extended thinking, yet it ended up with a better position. The game lasted 5 hours and was still unfinished.