AJ

[AJ @AjAbsaki]

@antirez any suggestions on bumping up the speed? max I’m getting is 3tps gen. this is really awesome work

[AJ @AjAbsaki]

@antirez wow - awesome job!!

[AJ @AjAbsaki]

@Howaboua this looks really cool.. unfortunately the install stalls on Mac.. hope to try it out in the future

[AJ @AjAbsaki]

@NousResearch mac app seems to be broken. thinks its in some sort of sandbox cant call tools to operate on my machine. telegram works fine, but the desktop app isnt working. have updated to the latest. will wait for a fix!

[AJ @AjAbsaki]

@The_Kyle_Mann @TheBabylonBee 😂

[AJ @AjAbsaki]

@ivanfioravanti as always - awesome work mate!

[AJ @AjAbsaki]

@noriqos @de__gods @nikitabier @ABernNYC @disclosetv @SpaceflightNow LOL

[AJ @AjAbsaki]

erm why does the deaktop app (mac) insist on downloading all models from hf? it started with just a tiny qwen and carries on downloading larger and larger models in the background. had to delete the app as there was no stopping it. Also does not work with openai compatible endpoints (no LM Studio or custom servers) - love the idea but the app is not usable for local models on a mac at this point. hopefully youll fix these basic issues soon would love to try again

[AJ @AjAbsaki]

@antirez what about the new step 3.7 flash? may be a good option to test

[AJ @AjAbsaki]

@zeddotdev very cool. however, the terminal sessions disappear on restart - so not sure how useful it is. had a grok build terminal session going and it was gone once I restarted the app. pretty painful

[AJ @AjAbsaki]

@muse_inmotion @elonmusk yes

[AJ @AjAbsaki]

@carbac80 @elonmusk ROFL - retard

[AJ @AjAbsaki]

@ivanfioravanti awesome work man well done!

[AJ @AjAbsaki]

@ivanfioravanti @antirez @stevibe Q2 imatrix did >90% on all benchlocal tests for me - at 30-38tps in a 128GB m5 max

[AJ @AjAbsaki]

@u1tra_instinct @0xSero try this - https://t.co/2FwiyCUvHT works great on a 128 m5 pro

[antirez @antirez]

Welcome to DS4, a specialized inference engine for DeepSeek v4 Flash. https://t.co/UrUJz5I2R1

This project would have been impossible without the existence of llama.cpp and GGML and the work of @ggerganov and all the other contributors. Thanks!

[AJ @AjAbsaki]

you’ve done something really special with ds4 mate. I use benchlocal to test quality and what you built qith your q2 on a m5 max 128 blows everything else out of the water. zero OOM ultra long context with no speed degradation and 90% + across all quality test suites. Cant wait to see what you build next

[AJ @AjAbsaki]

@_Adam_Holland @elonmusk in this version eve thinks you’re a retard, adam.

[AJ @AjAbsaki]

Thats really good - but for local models’ quality testing i’ve found benchlocal to be a good holistic capabilities test. DS4-q2 kills it (>90% on all tests). MTPLX Qwen (quality tuned) doesnt do too well. Will take another look to see if the fails are due to timeouts or actual quality related issues.

[AJ @AjAbsaki]

@KobeissiLetter ah yes - purposefully displaying a lower timeframe snippet with no overall context - pure engagement farming - here's two thumbs up for you 👍👍