frostjack980
218 posts
wow y'all love 5.5
we should think of something nice to do to celebrate!
English
@NVIDIAAP @vllm_project yeah except its not working on NIM for 4 days...
English
📊 Day 0 performance is here: DeepSeek-V4-Pro running on NVIDIA Blackwell Ultra.
Using @vllm_project's Day 0 recipe, we’ve captured the initial performance Pareto for DeepSeek’s flagship 1M long-context model. This curve highlights the baseline for balancing AI factory throughput with real-time user interactivity.
This is just the starting line. Expect these numbers to climb as we optimize the full co-design stack, including:
• NVFP4 & Dynamo
• Optimized CUDA kernels
• Advanced parallelization techniques and beyond
Read the full technical deep dive: nvda.ws/4sYg0D4
English
@jahirsheikh8 "Why is nobody talking about this?"
Well yeah its pretty much unusable now because everyone is talking about it and thanks to openclaw nerds
English
Why is nobody talking about this?
NVIDIA is giving developers free access to ~80 hosted AI models via API.
Models include:
• MiniMax M2.7
• GLM 5.1
• Kimi 2.5
• DeepSeek 3.2
• GPT-OSS-120B
• Sarvam-M
…and more.
Works with Cursor, Zed, OpenCode, Hermes, OpenClaude, etc.
Setup:
1. Get API key → build.nvidia.com/models
2. Use base URL: "integrate.api.nvidia.com/v1"
3. Choose model ID (example: `minimaxai/minimax-m2.7`)
4. Plug into your AI IDE / app
If you're experimenting with AI apps,
this is basically free inference.
Most builders have no idea this exists.
Thank me later.
English
@teortaxesTex If only k2.6 didnt burn through tokens like theres no tomorrow... Wouldve been my main model
English
interesting claim from Moonshot
V4 has more pretraining tokens (much more text-only), more active params, and yet does not dethrone K2.6, indeed it is *3 points behind* on SWE-Pro (according to V4's own paper).
Does Kimi really just have better data (at least in some domains)?
Shuyao Tim Xu@TimXu222575
@teortaxesTex not the amount of (pre/post-training) data, but the quality (how many engineers cleaning data)
English
@marmaduke091 LITERALLY the same thing was said about gpt 5.3 but here we are. always the same bs
English
🎉Congratulations to the @MiniMax_AI team on the launch of MiniMax M2.7!
MiniMax M2.7 is now available with NVIDIA GPU accelerated endpoints ready to try out with claws including NemoClaw and @OpenClaw. 🦞
📝Get started with our technical guide: developer.nvidia.com/blog/minimax-m… and see how you can begin experimenting for free at build.nvidia.com/minimaxai/mini….
What will you build this weekend?
Share in comments. 👇
MiniMax (official)@MiniMax_AI
We're delighted to announce that MiniMax M2.7 is now officially open source. With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%). You can find it on Hugging Face now. Enjoy!🤗 huggingface：huggingface.co/MiniMaxAI/Mini… Blog: minimax.io/news/minimax-m… MiniMax API: platform.minimax.io
English
@sendmeloc @neosovietposter It moved back. I guess you don’t follow the news?
English
Going from “a whole civilization will die tonight” to “yeah we surrender” is truly magnificent lol
English
Regarding file upload:
an input of 235K tokens overfills DeepSeek-Expert's context by 76%. Yes, expert, not instant. Thus, its context length limit is ≈133K (let's say 128).
Is it actually just a V3.3 after all?!
Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex
Very, very strange release is brewing. - unlimited fast model, as I've said, is probably what we have now, a V4-lite that's smaller than V3.2, they want it to be Pareto-stronger + 1M context - Expert not having file uploads because of token costs makes little sense, prefill must be very cheap now, and they could reduce file size? I often upload 20-30kb snippets - Vision being segregated into its own class is also unusual (vision would help bigly with expert workloads). I know for a fact that they've rejected deploying DS-VL on web previously because it wasn't ready for prime time/"not a big enough deal", so I don't think DS-OCR (a work that's apparently even older and smaller-scale) is directly involved here, and they probably use DS-OCR for OCR in chat by default. This Vision must be indicative of a fully functional VLM.
English
@kimmonismus Same womp womp each year, same monkeys falling for it
English
Holy moly: Sam Altman told Axios in a half-hour interview that AI superintelligence is so close, so mind-bending, so disruptive that America needs a new social contract.
- It's on the scale of the Progressive Era in the early 1900s, and the New Deal during the Great Depression.
- Altman warns: widespread job loss, cyberattacks, social upheaval, machines man can't control
- "soon-to-be-released AI models could enable a world-shaking cyberattack this year. "I think that's totally possible," Altman said. "I suspect in the next year, we will see significant threats we have to mitigate from cyber."
Mike Allen@mikeallen
🚨🚨@sama tells me he feels such URGENCY about the power of coming AI models that @OpenAI is unveiling a New Deal for superintelligence - ideas to wake up DC He says AI will soon be so mindbending that we need a new social contract 👇Altman's top 6 ideas axios.com/2026/04/06/beh…
English
Google’s new Veo model also seems to be on the way, possibly Veo 4?
x.com/medhini_n/stat…
Medhini Narasimhan@medhini_n
English
@Aryan_Ali_Pro @kimmonismus "click-bait" not the first time, def not the last
English
Confirmed by Fortune: Opus 5: so good that it poses a danger.
It's a massive leap from version 4.6, and it's so advanced that it poses a huge security risk.
They'll roll it out slowly because they're concerned about unforeseen consequences, especially security risks.
Besides Anthropic, OpenAI has also made similar announcements. As The Information recently reported, their upcoming model, "Spud," is an equally massive leap that will greatly accelerate the economy.
In short, 2026 will be the year everything changes. OpenAI has already renamed its department "AGI Deployment."
Chubby♨️@kimmonismus
Holy sh*t huge leak Claude Opus 5 "Mythos" leak: Claude Mythos, a new tier beyond Opus models, delivering major performance gains compared to Opus 4.6 in coding, academic reasoning, and cybersecurity, and signaling a step-change in frontier AI capability. The model is so powerful (and compute-intensive) that access will roll out slowly, starting with select cybersecurity partners to prepare for AI-driven exploits. This isn’t just a model release, it’s a preview of a new class of systems that could outpace current defenses and reshape how we think about AI risk and deployment. Take it with a grain of salt, @M1Astra is a very relaible source!
English
best coding models right now
sync work, shorter iterations, 1-5 agents in parallel:
- plan: composer 2
- work: composer 2
when you're in control, you simply only need composer 2.
async work, longer iterations, deep work:
- work: gpt 5.4
for async, you still want maximum intelligence as speed is less important with a model that can run for a really really long time
ui & taste
- opus 4.6 is still really strong here
- composer 2 is a fast follow, but not as strong (yet)
this space moves fast, so it will probably change soon!
English
@sergioluisduque @developedbyed idk i used to think the same until i started trying flash more and I can say there are many cases where Gemini 3 Flash is beating the pro version ( instructions following, hallucinations, even small coding tasks etc..). Give it a shot
English
@frostjack972755 @developedbyed how big of an improvement can regular version be? i mean, its not like people will stop using PRO. Main usecase i think is for those using it via the API for chat bots,etc?
English
@TheVsk_ @Stc_GadgetHubNG You're cooked ma boi why comments
English
