← sources

How I tried to like DeepSeek-V4.1-Flash-1017 only to realize that it is Opus 5.5 for poor

source: reddit (reddit:LocalLLaMA)published 2026-10-23 17:46ingested 2026-10-24 02:40

retained text

I tried to like DeepSeek-V4.1-Flash-1017, but Qwen3.8-Flash-Next feels much better for me. Let me tell you three cases where DSF failed me.

The first one was funny. I asked DSF to optimize project-maya configuration. After all options were explored I asked about possible upgrades that I can make and DSF started to insist that my HP Z440 with E5 2696v4 has AVX-512 instruction set. I told it that this is not the case, but it didn't believe me and wrote Python script to check AVX-512 flags. No surprise they came False for all AVX-512 variants. At first I was happy that DSF takes everything with gran of salt, but it continued to insist that probably I switched of AVX-512 in the bios, because it was completely sure that Broadwell-EP architecture has AVX-512 instructions. I never had a problem like that with QFN.

The second one was sever. I asked DSF to download academic papers from libgen site and it flat out refused due to the copyright infringements. I told it that I just need to make factual check of reference in my literature review section, but no luck. After several rounds of exchanges I said f-word to DSF and said that I am not going to be lectured by a model that was trained on the whole Internet without anyone's consent regarding paywall of f-g academic journals that take money for articles that were written on public money and who use unpaid labor of editors. QFN completed task in 20m without any questions.

The third is just matter of taste, but I enjoyed RP with QFN much more.

After that I told to myself may be it will be more intelligent in research. And decided to give it last chance and to pit all three best local models that I can run. The third model was GLM5.3-Flash. I asked all of them to compare current hypothetical AI bubble to the British Railway Mania (1840s), the Dot-Com Bubble (1995-2002) and to make a forecast. I am not going to lie DSF report looked very nice. It had nice figures and solid grounding. QFN report was not bad, but presentation was much better from DSF. Report from GLM was the shortest, with worst formatting and lacked any numerical analysis. At that moment I thought that it would be nice to have a reference point from a SOTA model and decided to add Opus 5.5 to the picture. It is here where things got weird. I looked at Opus 5.5 report and had a flashback of DSF report in everything: formatting, analysis, presentation. So I asked Opus to review all four reports and to say which are the most two similar. Here are Opus findings, D for DSF, Q for QFN, G for GLM5.3-Flash and O for Opus 5.5:

    ┌──────┬──────────────────────────┬───────────────────────┬────────────────┐
    │ Pair │ Word similarity (TF-IDF) │ Shared 4-word phrases │ Shared numbers │
    ├──────┼──────────────────────────┼───────────────────────┼────────────────┤
    │ D–O  │ 0.572                    │ 200                   │ 0.207          │
    ├──────┼──────────────────────────┼───────────────────────┼────────────────┤
    │ O–Q  │ 0.497                    │ 63                    │ 0.199          │
    ├──────┼──────────────────────────┼───────────────────────┼────────────────┤
    │ D–Q  │ 0.496                    │ 108                   │ 0.167          │
    ├──────┼──────────────────────────┼───────────────────────┼────────────────┤
    │ G–O  │ 0.469                    │ 57                    │ 0.156          │
    ├──────┼──────────────────────────┼───────────────────────┼────────────────┤
    │ D–G  │ 0.437                    │ 65                    │ 0.130          │
    ├──────┼──────────────────────────┼───────────────────────┼────────────────┤
    │ G–Q  │ 0.431                    │ 75                    │ 0.234          │
    └──────┴──────────────────────────┴───────────────────────┴────────────────┘

Look at the second column. I mean it is not just similar vocabulary it is straight 200 four word phrases. Concerning methods used in analysis they were also almost identical for DSF and Opus 5.5.

I really don't know what to think especially taking into account that Opus 5.5 is my favorite SOTA model, but somehow I don't like DSF results. What is you experience with these local "SOTA" models?

Bonus. We all are waiting for downfall of GPU and RAM prices, so here is a table with forecast from all 4 models. Hold on my local brothers the end is near.

https://preview.redd.it/7ent4csz69xh1.png?width=1930&format=png&auto=webp&s=38fe2baaab09c431ce740db59aa6eb68e8fa8a19

document fetched 2026-10-24 10:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xdkvmi/how_i_tried_to_like_deepseekv41flash1017_only_to