Erman Eroğlu

Mahallesiz/Klansız - Öteki (Yoğurdu Ekşi Olan), Software Developer (Namıdiğer Yazılımcı) - AI enthusiast 🤖

[Erman Eroğlu @geldeki]

@ivanfioravanti 🥳

[Erman Eroğlu @geldeki]

@PolymarketMoney 😂😂

[Erman Eroğlu @geldeki]

@fpostigo @alexocheema @TheAhmadOsman @AlicanKiraz0 @spark_arena @NVIDIAAI @NVIDIARTXSpark @SpaceTimeViking @grok tell me if aeon repo have something related to my 2xRTX 3090 rig

[Erman Eroğlu @geldeki]

@foley2k2 @TeksEdge Absolutely but even for a brand new device they should have thought these things. At least numbers should be better.

[Erman Eroğlu @geldeki]

Qwopus3.6 27B Coder MTP GGUF — llama.cpp draft-mtp results 🚀
Runtime: llama.cpp + CUDA
Quant: Q4_K_M
Context: 128K / 262K
Format: GGUF
MTP: enabled via --spec-type draft-mtp
1× RTX 3090, 128K context:
• Short coding → ~22.6 tok/s
• Medium architecture → ~25.4 tok/s
• Pure coding → ~22.7 tok/s
• Before MTP: ~19 tok/s
2× RTX 3090, 262K context:
• Short coding → ~40.3 tok/s
• Medium architecture → ~47.6 tok/s
• Pure coding → ~42.7 tok/s
• Long prompt test → ~40.1 gen tok/s
• 258K prompt stress test passed earlier
Takeaway:
MTP makes Qwopus much more interesting locally.
Single 3090 can run 128K nicely, while dual 3090 makes near-native 262K context feel genuinely usable.

[Erman Eroğlu @geldeki]

Stress-tested Qwopus3.6 27B Coder MTP GGUF near native 262K context on my local 2× RTX 3090 rig 🚀
Runtime: llama.cpp + CUDA
Quant: Q4_K_M
Context: 262K server, native metadata
Hardware: 2× RTX 3090 24GB
Results:
221K prompt tokens → 305.5 prompt tok/s, 18.35 gen tok/s
251K prompt tokens → 226.4 prompt tok/s, 17.62 gen tok/s
258K prompt tokens → 216.9 prompt tok/s, 17.45 gen tok/s
Also tested 32K daily mode at ~26 tok/s stable generation.

[Erman Eroğlu @geldeki]

@TeksEdge I think it is a fair price but I wish it was less though

[Erman Eroğlu @geldeki]

@arafatkatze US is feeding its own enemies by their politics :

[Anthropic @AnthropicAI]

The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.

The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.

Access to all other Claude models is not affected.

We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.

Read our full statement: https://t.co/bwn0sximKZ

[Erman Eroğlu @geldeki]

@SinanTankutGlhn Şu Avrupa neden yapay zekada bu Çin ve ABD nin gerisinde kaldı anlamak mümkün değil. Acısını çekerler böyle giderse ve umarım Anthropic in Avrupa’ya veya Asya ülkelerine taşınma planları vardır.

[Erman Eroğlu @geldeki]

@AMD While cuda is there and with a higher higher bandwidth are there any reasons to choose AMD over Nvidia DGX Soark at same prices?

[Erman Eroğlu @geldeki]

@DavideTortoise @AMD Even with less bandwidth I guess

[Erman Eroğlu @geldeki]

@KyleHessling1 Yeah I keep trying thanks!

[Erman Eroğlu @geldeki]

@KyleHessling1 Can be better with some adjustments

[Erman Eroğlu @geldeki]

@KyleHessling1 Got Qwopus3.6 27B Coder MTP GGUF running locally on 2× RTX 3090 🚀

llama.cpp + CUDA
Q4_K_M
32K context
-b 4096 / -ub 1024
--split-mode row
--jinja

Result: ~26 tok/s stable generation.

[Erman Eroğlu @geldeki]

Dual RTX 3090 benchmark:
Qwen 3.6 27B AutoRound INT4
vLLM, dual GPU, 262K max context
Generation speed:
• Avg: ~91.8 tok/s
• Max: ~95.2 tok/s
• 5 runs × 1000 output tokens
Long-context:
• 195K prompt tokens processed successfully
• 255K prompt tokens processed successfully
• 255K prompt end-to-end time: ~89.9s
Dual RTX 3090s are still a very capable local LLM setup.
#LocalAI #LLM #RTX3090 #Qwen #vLLM

[Erman Eroğlu @geldeki]

Can’t wait benchmarking this with my dual RTX 3090s 🥳

[Kyle Hessling @KyleHessling1]

Qwopus 3.6 27b-Coder is now live!

Scores a 67% on a full run of SWE bench verified with thinking completely disabled! Q5_K_M

This model is lightning fast for dense class! With a natively finetuned MTP head, it achieves 100 tps on a single 5090! The biggest upgrade here, though, is its stability in programming and tool calling within @NousResearch Hermes agent, with thinking off!

Wall time is crazy fast this way, which makes Hermes feel "native" and snappy, like they were meant for each other. The freedom of running without thinking at all makes you part of the thinking process, and you never get caught waiting 15 minutes for it to finish a thought string, like with the base models.

Thinking on and temp high, .9-1 seems to produce really incredible design and svg results. I reran the Boat survival prompt through a few turns, thinking on, and it seemed to render more fancy models in HTML canvas, but it was much more of a start-a-prompt and wait experience vs the snappy and active iteration with it disabled. It may be worth turning it off and on throughout the build process if you want to get really creative with design.

Really looking forward to seeing how this one performs for y'all! Please post comments with your opinions and use cases below! As always with our fine-tunes, mess with the temperature setting, and run them much hotter than the base!

Please check out the Boat Survival game I posted yesterday, made in 12 turns using Hermes and this model, with thinking off. Link below!

Full swe bench repo-specific breakdown also posted in the comments for those interested!

Happy building, everyone! We're looking forward to your thoughts! Quants uploading now!

https://t.co/kxJE3C39ZZ

[Kyle Hessling @KyleHessling1]

Qwopus 3.6 27b-Coder is now live!

Scores a 67% on a full run of SWE bench verified with thinking completely disabled! Q5_K_M

This model is lightning fast for dense class! With a natively finetuned MTP head, it achieves 100 tps on a single 5090! The biggest upgrade here, though, is its stability in programming and tool calling within @NousResearch Hermes agent, with thinking off!

Wall time is crazy fast this way, which makes Hermes feel "native" and snappy, like they were meant for each other. The freedom of running without thinking at all makes you part of the thinking process, and you never get caught waiting 15 minutes for it to finish a thought string, like with the base models.

Thinking on and temp high, .9-1 seems to produce really incredible design and svg results. I reran the Boat survival prompt through a few turns, thinking on, and it seemed to render more fancy models in HTML canvas, but it was much more of a start-a-prompt and wait experience vs the snappy and active iteration with it disabled. It may be worth turning it off and on throughout the build process if you want to get really creative with design.

Really looking forward to seeing how this one performs for y'all! Please post comments with your opinions and use cases below! As always with our fine-tunes, mess with the temperature setting, and run them much hotter than the base!

Please check out the Boat Survival game I posted yesterday, made in 12 turns using Hermes and this model, with thinking off. Link below!

Full swe bench repo-specific breakdown also posted in the comments for those interested!

Happy building, everyone! We're looking forward to your thoughts! Quants uploading now!

https://t.co/kxJE3C39ZZ

[Erman Eroğlu @geldeki]

@Youssofal_ 🥳

[Erman Eroğlu @geldeki]

@UnslothAI 26*

[Erman Eroğlu @geldeki]

@UnslothAI Btw guys post is for 25 6b. I posted under 12b