Christian Balbin

Qwen3.6 27B and 35B-A3B are amazing models, but nothing reaches the efficiency of GPT-OSS yet. Qwen3.6 35B-A3B is as fast as GPT-OSS-20B but nowhere near the prefill performance.

@OpenAIDevs @coreyching on the toilet seat is wild

goalmaxxing

@jxnlco as if you don’t have like 50 MacBooks 😆

People don't know this, but I actually leave my laptop at the office on the weekends, and I skate in every time I need to check my computer.

@TheZachMueller want

In today’s adventures of working at Lambda is super cool, I get to have a B200 rack in my house*

@BonannoDavid fine. I’ll follow

I’m not doing this X thing correctly. I’m the CFO of a company that just announced the largest crypto M&A deal in history, I posted about it, then reposted my own posts and somehow I only gained one follower this week…. Can someone tell me what I’m doing wrong?

@bnjmn_marie Your account is literally straight alpha for local LLMs. Thanks for doing this !

I benchmarked Google’s new MTP for Gemma 4 31B using vLLM with 4 speculative tokens, a fairly conservative setup. Results: - Much higher throughput than Qwen3.6’s MTP - Lower latency too, helped by Gemma 4 generating fewer tokens - For coding tasks with reasoning enabled, Gemma 4 is now at least 6x faster than Qwen3.6. So you can generate 5 outputs, run your tests to select the best one, and it would still be cheaper than a single output by Qwen3.6. I’ve updated my full comparison with the new numbers: kaitchup.substack.com/p/qwen36-27b-v… I also confirmed what others have reported: Gemma 4’s MTP handles a high number of speculative tokens very well. On simple text generation, I’m now testing values above 10 and reached 129 tok/s on an RTX Pro 6000, compared with 20 tok/s without MTP. Next step: confirming how this translates to real tasks.

@loktar00 @vllm_project same here. also the rc broke the generation of xml/html in tool calls because of a parsing issue.

@vllm_project Bah even with nightly 20.2.rc1 I keep getting: NotImplementedError: Speculative Decoding with draft models or parallel drafting does not support multimodal models yet

🚀 Day-0 MTP support for Gemma4 now available at vLLM with ready-to-use docker image! ⚡️Enjoy up to 3x faster decoding performance to supercharge your development with zero quality degradation! Check out the full vLLM recipes for Gemma 4 model series👇 recipes.vllm.ai/Google/gemma-4…

@YouJiacheng they should know these are wild claims and that people would rightfully be skeptical. they should have had trusted a 3rd party benchmark them

81.8% swe-bench in the first release??? wow. is this legit?

@rileybrown i'll believe it when i see it on @ArtificialAnlys

Chat is this real?

@sama 👋

hey chat, we haven't forgotten about you 👀

@OpenAIDevs @thsottiaux @jxnlco are legends . That’s a lot of tokens…..

@jxnlco @jxmnop they did not. 1M context is still the default. I agree it feels like it degrades past 200k tho

@jxmnop They did?

it is endlessly fascinating to me that we still don't have a true 1M-context model it's an unusual case where the infra is far ahead of the science. Claude discontinued 1M+ context bc it didn't really work past ~200k we don't have the right data? training techniques? not sure

how i feel using xhigh to make the most simple UI cosmetic changes

@sama that 'not accepted' email spiked my cortisol lol but appreciate the kindness here. OpenAI is one of the few companies where it really feels like they care deeply about the community

we are gonna do something nice for everyone who applied for the GPT-5.5 party and that we didn't have space for. hope you enjoy!

anyone using ChatGPT Atlas ? just remembered about it when going through my apps

@daradoescode It's crazy that computer use has basically been solved

Computer use shenanigans

@dudufolio bro don’t make your agents work under pressure

when your agents are running but you have to reach the top shelf

@MatthewBerman i would just run the draft model on the dgx station. My assumption is you won’t get too much of a gain by offloading it to the spark.

Pretty sure I can double the t/s Isn’t possible to do speculative decoding with a DGX spark on the same network?