@xoofx @GergelyOrosz The announcement about Sign in with ChatGPT (with Pi, OpenCode etc in the initial list) makes me concerned OpenAI may move to no longer support it outside of the new method and for high profile open harnesses. Hopefully an unjustified worry.
1
0
1
129
Alex Bradbury retweeted
Landed fixes in GCC & Clang for [expr.static.cast]/8 wrong-code bugs when investigating a V8 bug related to it. Previously:
enum Flag : bool { kNo, kYes };
static_cast<Flag>(2); // kNo in GCC ❌, kYes in Clang ✅
static_cast<Flag>(0.5); // kNo in GCC and Clang ❌
enum Flag : bool { kNo, kYes };
static_cast<Flag>(2); // kNo in GCC ❌, kYes in Clang ✅
static_cast<Flag>(0.5); // kNo in GCC and Clang ❌
2
4
30
2,756
@ZixuanLi_ @huggingface Thank you for sticking to a standard open license. Amazing release!
0
0
0
320
@ctnzr Why "many of"? What would make you choose a different license for a Nemotron release?
1
0
0
531
@serenaa_ge This "failure modes by model" analysis is fantastic. Any plans to extend it to cover the tested open weight models?
0
0
1
661
@FireworksAI_HQ trust.fireworks.ai lists security@fireworks.ai at the top presumably as the contact for questions on the privacy policy, certifications, and subprocessors. One of my colleagues reached out and had no response for a couple of weeks now. Is the email monitored?
0
0
2
223
@_TobiasLee @natolambert @XiaomiMiMo @xeophon Wow I didn't have to wait for long. Both variants, MIT licensed, based models too. What a release!
0
0
3
24
@natolambert @_TobiasLee @XiaomiMiMo @xeophon Any word on whether both Mimo V2.5 and V2.5 Pro will be open weights, or just V2.5?
1
0
1
814
@xoofx I simply don't trust myself to consistently make the right decision with these constant requests for permission. Too easy to get lazy. Providing an environment thst you're happy to let the agent run in YOLO mode within seems the only sensible path to me.
0
0
0
132
@TheZachMueller @casper_hansen_ Until then I'll just hammer F5 until Kimi K2.5 appears!
0
0
1
32
@eliebakouch Does this report confirm one way or another if they got access to a non-public kimi-k2.5-base, or if they built on the instruction tuned open weights release?
0
0
0
268
@TheZachMueller @nvidia I love this model card effort. Is Kimi K2.5 on your to-do list?
1
0
1
109
@TheZachMueller @LambdaAPI @Zai_org @arcee_ai I wrote a short post that converts your reported results to the implied Wh per query (with many assumptions!) muxup.com/2026q1/minipos…
0
0
0
43
@TheZachMueller @LambdaAPI @Zai_org @arcee_ai The ITL doesn't seem to match what you'd need for 700tok/s with up to 32 parallel requests, but perhaps it's measuring the latency between groups of tokens returned. Other models you benchmark don't have this mismatch but also aren't using speculative decoding
1
0
0
48
@TheZachMueller @LambdaAPI @Zai_org @arcee_ai This is great! It looks like there's an error on the glm-5 page as the table lists total tokens in and total tokens out corresponding to 512 8k/1k prompts while the rest of the page suggests 256 were used.
1
0
0
65
@natolambert For slide 17 on the eval talk "20% of Olmo 3 post training compute was for evals": I understand pretraining for the two base models was ~1.24M GPU hours. Is this slide really saying >1.4e6 (1.4M) GPU hours for post training or am I interpreting it incorrectly?
0
0
0
80