New 0731 iq2 thinks forever!

[arkham000]

Hi, I downloaded DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2-imatrix-0731.gguf from antirez HF repo, it works quite well aside the facts that it thinks forever.
I'm using it through pi agent using the exact configuration from antirez which I previously used for the preview version, but thinking now has gone mad, I currently set it to low (but I seen no difference from medium or high) and to just update a bunch of files on my llm-wiki it generated 9k tokens of reasoning!
It goes all the the time Mhh wait let me reconsider... it became qwen3.5 all of a sudden lol

Have you observed the same problem? Are there any workaround ?
Like this it is unusable, I'm on a m2 max with 96GB and between the low inference speed and infinite thinking it is very painful to use ....
Thanks

[TipKnuckle]

Counterintuitively you should try setting the reasoning to 'max'. I can't find it now, but on one of the LLM subreddits there was thread were a user did some testing and found this version of DeepSeek Flash 4 did a ton more reasoning in 'low' compared to 'max.' The model is working well in my experience over the past few days.

[arkham000]

Yeah I found that post thanks. I'll try it

…

On Mon, 3 Aug 2026 at 22:52, Craig ***@***.***> wrote: *TipKnuckle* left a comment (antirez/ds4#660) <#660 (comment)> Counterintuitively you should try setting the reasoning to 'max'. I can't find it now, but on one of the LLM subreddits there was thread were a user did some testing and found this version of DeepSeek Flash 4 did a ton more reasoning in 'low' compared to 'max.' The model is working well in my experience over the past few days. — Reply to this email directly, view it on GitHub <#660?email_source=notifications&email_token=BG35SQWETNJJ63E6GCB2FST5ID3SVA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMJXGE2TQOJRGQY2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KYZTPN52GK4S7MNWGSY3L#issuecomment-5171589141>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BG35SQQMXR26VNY3A2AHZVD5ID3SVAVCNFSNUABGKJSXA33TNF2G64TZHMYTEMZRGE2DQNRSGY5US43TOVSTWNJQGQ2DQOBWGEZTTILWAI> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/BG35SQWRWJWKKOG2LPUBOHD5ID3SVA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMJXGE2TQOJRGQY2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2KUZTPN52GK4S7NFXXG> and Android <https://github.com/notifications/mobile/android/BG35SQTXM3Y2KMWEWE32O2D5ID3SVA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTKMJXGE2TQOJRGQY2M4TFMFZW63VGMF2XI2DPOKSWK5TFNZ2K4ZTPN52GK4S7MFXGI4TPNFSA>. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>

[Flor1an-B]

Found the likely root cause while looking into reasoning_effort more broadly (I have an open PR — #704 — for a related but separate bug in the "high"/"max" tiers).

Checked against DeepSeek-V4's own encoding spec (encoding/README.md): it defines exactly three reasoning_effort levels, and "low" is the default with no prompt prefix at all — it's meant to be the model's normal, unprompted reasoning behavior. But DS4's parser currently maps "low" (along with "medium"/"minimal") to the same internal tier as an explicit "high" request — the most elaborate non-max prompt DS4 has ("You MUST be very thorough... rigorously stress-testing your logic against all potential paths, edge cases, and adversarial scenarios..."). So setting reasoning_effort to "low" was actually asking the model to be maximally thorough, which is a very plausible explanation for the runaway reasoning you saw — not a quant-specific quirk.

(The "max sometimes reasons less than low" observation upthread lines up with this too: before my other PR's fix, "max" was also misrendering as the "high" text, so both settings were likely producing near-identical prompts either way.)

Pushed a fix as an additional commit on #704 that adds a real "low" tier (thinking enabled, but genuinely no extra prompt text, matching the spec). Would be good if you could confirm whether it resolves what you're seeing once it lands — happy to also test against the specific IQ2 quant you mentioned if that turns out to matter.