ggcr

I was also surprised to not find any benchmark for RTL in the model card. Wonder how it would score on - say, CVDP

From what I've read, it's the first time a (C++) draft hasn't been passed by consensus. So what happens now?

Long-horizon RL symptoms perhaps?

Comic Code ;) You won't regret it, been using it 5 years.

https://tosche.net/fonts/comic-code

Meta's blog post still list it as 2nd in AA. Any ways, AA moves benches constantly, when Qwen3.8 was released something similar happened.

RL team hill-climbing this to 2nd AA place and this fast is crazy

How does it work on day-to-day workflows? does it translates well?

Oh yeah totally, good point. I remember there was controversy around Claude tokenizer generating more tokens lol

What's impeding a lab from releasing its own new model, pricing it really low for the beautiful Pareto plot, accompanied by phrases VC love like "establishing a new frontier in cost", to just then raise prices back up?

Live notification in Codex:

> GPT-5.6-Sol is retiring. This conversation will automatically switch to GPT-6-Sol

I don't recall OAI retiring a model so early lol. Similar arch?

Interesting. Perhaps I can see this being quickly adopted in LLMs-as-a-judge, where you normally need (a) a structured answer, say, with lots of different fields (metrics) and (b) you want the judge to be fast, not being a bottleneck.

The CTO himself reported 600 commits and 300 PRs each day. And they can do so because they have a massive CI [1]:

> Every day CI runs about 20..80 million tests in 600 commits and 300 pull requests

> Last year, ClickHouse spent 360 years of machine time for CI

I am no user of CH so I can't talk about their product. But we are talking about a company with 686 employees as per their LinkedIn, where ClickHouse is clearly the core of their business. Considering all of this, is 50+ commits a day that much?

[1] https://presentations.clickhouse.com/2026-openhouse-sf/great...

> I am so used to using one hand to operate phones.

I even forgot about this. God damn it, I enjoyed the SE so much because of this.

Seems like a big engineering feat by Apple, but for real, who asked for this?

Edit: Is there that many demand for a foldable iPhone that justifies Apple making such engineering investment?

Reminds me, kinda, to when Astra was launched and OpenAI announced an improvement to the bounded prime gap. Which BTW, Prof. Julia Stadlmann had published an independent result only a few days earlier

Stadlmann improved it from 246 to 240, OpenAI later claimed 186 I think?

Maybe someone can help clarify? I am no expert at all, but I can't help but see similarities.

[0] https://arxiv.org/abs/2608.31126

> Let me make plain what I have said to colleagues in private: in view of this body of work, I believe Luis Martínez-Zoroa deserves a Fields Medal.

This would go hard if I was an OpenAI agent

2026 has been the year where spec-dec has matured, it has been adopted by all big OSS engines and i'm sure it's present in quite a lot of inference providers as the default

I feel like P/D dissaggregation will be the next big one for providers, as prefill tends to be compute bound while decode mem bound which I guess each will have a different type of node

Here is an employee of HF at a time answering how do they make money, supposedly [1]

> We make money via compute credits + Enterprise Hub + HF Pro subs [...]

I guess this plus custom inference deployments, external inference providers, partnerships with the big cloud AWS, Azure, etc.

[1] https://x.com/reach_vb/status/1928050126498713706

Agree. I'm not into interpretability but I always try to follow what Bau and his team are up to.

Z.ai founders were one of the pioneers in MM-LLMs with Cog-VLM years ago, back when LLaVA emerged. I wouldn't be surprised if they added multi-modal capabilities

Ziphu has that many resources to be able to serve capacity for 1 quadrillion tokens per day on Nous portal? My bet is that it's a Composer model from Cursor running on xAI cluster, they already did a Composer based on Kimi-K2.5

Announcement tweet by Stanford prof Percy Liang and Marin: https://x.com/percyliang/status/2090918065634684997 https://xcancel.com/percyliang/status/2090918065634684997

It will include Pre-training and Mid-training. ETA 3 months.

Live wandb dashboard: https://wandb.ai/marin-community/marin_moe/reports/535B-A23B...

GitHub issue tracking progress: https://github.com/marin-community/marin/issues/8435

23T tokens (not deduped) directly in S3 for anyone to see: s3://marin-us-east-02a/marin/datakit/store_4d2e363d

Intermediate checkpoints also in S3

They should've included Qwen3.5-397B-A17B in the benches then :/

TIL that alphabet.com belongs to BMW group

> nvda will have an incentive to continue this kind of releases, even if other parties slowly abandon the open release of models

This! It's literally in their best interest for open-weights models to succeed

Nice cadence of releases by the Nemotron team :)

> Oriol Vinyals, Sanjay Ghemawat, Jeff Dean, Quoc Le

as founding members is crazy !

Blog: https://clickhouse.com/blog/andy-pavlo-joins-clickhouse?utm_...

Waiting for Qwen3.8-27B :)

Their base models and architecture has quickly become the go-to for local inference and fine-tuning, even when they introduced some tricky things like GDN, so many people use it, that it was matter of days/weeks until lots of OSS frameworks adopted it.