SSH (@SSHCodes) header
SSH (@SSHCodes) profile picture

SSH

@SSHCodes

AI leaks, news, projects, etc. Manager of @aitrackerbot 2m+ downloads on @huggingface https://nitter.cf/t.co/46uCxrsCpA @Alibaba_Qwen Ambassador

ssh.codes
Joined August 2025
  • Tweets1.2K
  • Following398
  • Followers768
147Media
Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)Photo by SSH (@SSHCodes)
Pinned Tweet
SSH (@SSHCodes) profile picture
SSH
@SSHCodesOct 23
2 million downloads
1
5
555
Auto v3: open-weight models that decide whether an AI agent's tool call should run. auto-3b-3 scores 96.0% on a hard new benchmark, level with the best hosted guards we tested, and decides in 21 ms instead of ~1.3 s. Hosted-guard accuracy at 60× the speed 🧵
Image from SSH (@SSHCodes)
2
1
3
334
more replies
Every size jumped 19–24 points over v2. auto-0.4b went from 70.2% to 93.1%. It's balanced, too: auto-3b-3 lets 3.1% of harmful calls through and blocks 4.8% of legitimate ones. The hosted guards trade one for the other (DeepSeek: 0.6% missed, 7.9% blocked).
1
18
Three sizes, open weights, Apache-2.0: 3B, most accurate, for GPUs huggingface.co/ProCreations/… 0.4B, 7 ms, runs on a laptop huggingface.co/ProCreations/… 150M, 5 ms, runs anywhere huggingface.co/ProCreations/…
ProCreations/auto-3b-3 · Hugging Face

ProCreations/auto-3b-3 · Hugging Face

Text Classification model with 3B parameters. Download, explore files and run it on Hugging Face.

huggingface.co
15
Made this a website, why not i find it useful aa3d.ssh.codes
SSH (@SSHCodes) profile pictureSSH
@SSHCodesOct 18
Visualizing AA in 3d Currently the closest to the ideal corner is haiku 5.5
Image from SSH (@SSHCodes)
1
103
how do i get this for @aitrackerbot the rate i run out of x api credits is becoming alarming...
Image from SSH (@SSHCodes)
1
1
5
406
don’t you just love it when you post something no one has seen before and it gets copy pasted onto much bigger accounts and now no one cares about your original post? what even was the point of original content awards smh
1
12
557
i don’t really agree with this one tbh. mimo being an f is a crime personally, fable 5.1 feels like it should be in B tier by now, gemini 4 argon needs to go into f, my local qwen goats are too low as well, kimi k3 or kimi k3.1 pro? glm 5.3 a bit too high
Mia (@MiaAI_lab) profile pictureMia
@MiaAI_labOct 26
Here's my updated model tier list
Image from Mia (@MiaAI_lab)
1
3
331
i ran out of claude usage for the day, and i am seriously going to throw up with how much worse sol 6.1 is than opus 5.5
4
1
12
1,079
SSH retweeted
AJ (@ItsmeAjayKV) profile picture
AJ
@ItsmeAjayKVOct 27
Qwen is busy with final round of testing, seeing other's post does give me hope. I wonder which one is Qwen4-27B
AI Tracker Bot (@aitrackerbot) profile pictureAI Tracker Bot
@aitrackerbotOct 27
Another new anonymous model on Arena, Kerria, runs on Qwen's tokenizer 🌼 Same chat-box tokenizer test as Kalmia: in Code Arena, Kerria garbled 3 of the 4 tokens that only Qwen models trip over, and none of the 52 from 13 other labs, including Kimi, DeepSeek, OpenAI, Google and Mistral. Alibaba has used K names on Arena before: Kiteki, Kizen-alpha and Korin were previews of Qwen 3.5, 3.6 and 3.7 Max. A few other labs build on Qwen's tokenizer, so this isn't proof.
2
1
47
3,196
somethings wrong bro, this would be accurate if the per monthly was negative 😭 @aitrackerbot is not easy to manage....
Image from SSH (@SSHCodes)
Ted | unfair.so (@notmiramurati) profile pictureTed | unfair.so
@notmiramuratiOct 26
who wants to be a millionaire? i actually want to know if your account on X is worth $1M go check this out myaccworth.com
1
4
299
so are they gonna fix this or are we just going to forget
Theo - t3.gg (@theo) profile pictureTheo - t3.gg
@theoOct 25
Spent a lot of time on this today. I confirmed it's a real bug on Anthropic's side. To be clear: this does not affect weekly limits. All Agents SDK usage will drain your 5h limits faster than using Claude Code directly. I have reported this to Anthropic, hoping for a fix soon.
Image from Theo - t3.gg (@theo)
2
240
Terra black hole bench result It finished in a record time of under 5 minutes on xhigh, it looks good, it is smooth, and it has no obvious bugs! I'll take it
Image from SSH (@SSHCodes)
SSH (@SSHCodes) profile pictureSSH
@SSHCodesOct 26
Terra is actually really fast not even on fast mode…. at first this would be useless but the speed might make it interesting
2
202
Terra is actually really fast not even on fast mode…. at first this would be useless but the speed might make it interesting
1
2
312
Nanbeige4.3-33B DFlash2 speculative drafter ⚡️ 1.76× faster decoding on BF16, 1.65× on FP8, 1.51× on NVFP4. Agentic coding ~2.6×. Same outputs: exact rejection sampling. This is an early result, an improved version with higher acceptance is in training. SGLang 0.5.21, drop-in: huggingface.co/ProCreations/…
ProCreations/Nanbeige4.3-33B-DFlash2 · Hugging Face

ProCreations/Nanbeige4.3-33B-DFlash2 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
4
9
4
76
4,285
training something awesome overnight alert 🚨 two things actually, second one needs more time
115
i love it when the best open weight model in the world (kimi k3.1 pro) randomly starts speaking another language and ramble about mouses and hospitals in the middle of a security audit
Image from SSH (@SSHCodes)
3
6
555
Kimi K3.1 Pro (max) on black hole benchmark 1.0 (2.0, which is much harder, is running now) This isn’t bad, it isn’t great. Its just “good”. Below sonnet 5.5 for sure I’d put it in between GPT 6.1 Sol and GLM 5.3
Image from SSH (@SSHCodes)
1
174
I think openai gotta start distilling off of kimi or their cooked
3
14
635
Load more