SSH

I havent tested this new model myself but the AA scores look really good for its size, makes it even more exciting because this is just an intermediate checkpoint. pretty cool

If AI tracker bot detects it then the leak is true

@Lentils80 ...where?

🚨 Breaking: Gemini 3.6 Flash, ID "gemini-3.6-flash-tiered", appeared in Antigravity a few minutes ago

@aitrackerbot hello my friend

Hello, world.

Go follow my new bot, AI tracker! Will give you the latest leaks / model releases as fast as possible x.com/aitrackerbot?s…

how does a model you can run on a phone beat the latest haiku model from anthropic by a lot... haiku SUCKS

guys i have done the ultimate test of qwen 3.8 nothing.ssh.codes

@jawrooo_ true, and qwen's colors are nice

@SSHTheDev Idk I feel like this has a clean lighting effect which is hard to do. It’s not 100% realistic but imo it’s right behind fable which I think has the best

Does it beat gpt 5.6 sol? Kinda... personally I still like sol's better. #viewer" target="_blank" rel="nofollow noopener">bhbench.ssh.codes/model/qwen38#v…

@OmedVibeCodes Theres a secret limited time 10x promotion on their website for the token plans, kinda sketchy, makes me think as qwen 3.8 releases they want people to think limits are awesome then suddenly turn it off docs.qwencloud.com/token-plan/per…

Qwen’s Token Plan might be the most affordable coding plan right now at just $18. It’s basically Codex’s $100 plan, but around 5x cheaper. You get a ton of usage with Qwen 3.8 Max, which is pretty close to GPT-5.6 Sol. I ran it for 1.5 hours straight with constant outputs and still didn’t even hit 10% of the 5-hour limit. Kinda insane.

@basaltlabs Lmao ok you got us youtu.be/enk4w5mRjQY?si…

@basaltlabs Buddy you got some explaining to do

Introducing Monolith-1.0: Frontier Reasoning, Open Weights - 1.57T-param Mixture-of-Experts, 49.5B active per token - Native 1M-token context (2²⁰), extended via a two-stage YaRN curriculum - Trained on 60T tokens across 12,288 Ascend 910C NPUs - Frontier SOTA open-source model as of July 2026: 95.9% GPQA Diamond, 96.2% MMLU-Pro, 90%+ AIME 2025 - MIT-licensed weights, tokenizer, and eval harness No waitlist - open the chat and give it something that should be hard. Chat: basaltlabs.org/chat

@basaltlabs Also for anyone who doesn’t believe me, try the model yourself, it’s answers are genuinely garbage

“We need answer who are you”

thanks tibo

kimi K3 on BHBench. Not bad i guess, but some downsides. #viewer" target="_blank" rel="nofollow noopener">bhbench.ssh.codes/?m=kimi#viewer This was the most expensive open weight model run on BHbench yet, the model overthinked a ton and the accretion disk has tiles.

why is my codex app suddenly chatgpt app icon 😭 @thsottiaux we want codex not chatgpt

so… grug 35b has a pretty big error but I have burned a lot of GPU credits already and don’t want to use more. ideas?

what i love about gpt 5.6 sol: - it can work for hours upon hours doing real work, and by the end the work WILL be done (very determined) - it doesn't always need to work for hours though, at tasks that take less time it still does very well - It is fast compared to other models it is my new default for 100 percent of work (my fable usage somehow dropped to 0)

putting into comprehension how inefficient ultra mode is on codex rn: two days ago, i had gpt 5.6 sol xhigh literally running the ENTIRE day almost (18 hours nonstop). It used weekly 15 percent usage and 1b tokens Yesterday I used ultra mode just a bit (maybe 3 prompts) that used TWO BILLION TOKENS and 35 percent weekly usage even when I barely used any prompts and it worked for at most two hours