Aloha!🌺 Meet Ornith-1.5-122B: No Tradeoff Between Strong and Fast Stronger. Faster. Leaner. Ornith-1.5-122B delivers frontier-level capabilities with exceptional inference efficiency. With superior instruction following, stronger task decomposition, and deeper understanding of real-world workflows, it achieves state-of-the-art performance among open-source models of comparable size, while running on a single NVIDIA DGX card!! Open weights under the MIT license. Coding ✅SWE-bench Verified: 0.845 ✅TB2.1: 0.813 ✅CyberGym: 0.800 ✅NL2Repo: 0.512 Agentic ✅HLE: 0.446 ✅BrowseComp: 0.785 ✅MCP-Atlas: 0.746 ✅Toolathlon: 0.708 🤗Hugging Face: huggingface.co/ornith-ai/Orn… We can’t wait to see what you build with Ornith-1.5-122B.

Run DeepSeek V4.1 Flash 1017 on 2x DGX Sparks ✨ - 1M total KV pool, 512K context - 75 tok/s on prose, single stream - 130 tok/s on prose, 4 concurrent streams - Prefill 2000-2175 tok/s - EXL3-EAQv1, 2.9 bpw The quant in this recipe got about the same calibration as my GLM 5.3 Flash, with some extras. Feedback would be appreciated! Special thanks to @bertholomusai and all the contributors for the engine work, and to @ashxhart for TensorFold. Get it here: github.com/MiaAI-Lab/DeepSee…

one day you wrote console.log("here") for the last time and didn't even know it

The team at @Beelinkofficial sure knew how to make me feel welcome. Excited to announce the fruits of this visit soon!

Run DeepSeek V4.1 Flash 1017 on 2x DGX Sparks ✨ - 1M total KV pool, 512K context - 75 tok/s on prose, single stream - 130 tok/s on prose, 4 concurrent streams - Prefill 2000-2175 tok/s - EXL3-EAQv1, 2.9 bpw The quant in this recipe got about the same calibration as my GLM 5.3 Flash, with some extras. Feedback would be appreciated! Special thanks to @bertholomusai and all the contributors for the engine work, and to @ashxhart for TensorFold. Get it here: github.com/MiaAI-Lab/DeepSee…

Muse Spark Open -- weights releasing today -- runs great on new Mac Studios thanks to work with @johnternus and Apple.

1/ muse spark open is the most capable US open-weight model to date. 816B total params, 42B active. 1M-token context. natively multimodal. built for coding and agentic work like long-horizon tasks, hard software engineering, multi-step reasoning. america! 🇺🇲

Can I just say how tired I am of these kinds of remarks? They're getting more and more common as the internet has become armchair experts on all software and businesses. We build best in class software and make it available completly free. No one else in our category does that. No one. Don't want it? Dont use it. Sentry will absolutely run just fine on 32gb of memory, but thats not even the point. Its complex. That complexity comes from nearly two decades of development. Some of it is avoidable, some of it isnt. This is why cloud services exist - because we can amoratize the cost and complexity vs spending every waking hour optimizing every small part of a system to run on random low powered VPS. "Rewrite in Rust" - you don't maintain the software, nor do you undrestand what that means. An important chunk of Sentry is already in Rust (which could be C++ or Go!). The rest makes absolutely no sense to even suggest that. No one wants to work on a user facing web service with Rust, waiting on compile times non-stop. Nor do our customers. They want a great product, and they want to keep seeing improvements to that product. Our time is spent making it actually better for our 100k customers, not optimizing for your single use case. There is always opportunity for improvement, but takes like this are both ignorant and entitled.

ps5 can run steam games for FREE now windows games on a jailbroken ps5 here: github.com/leolobato/playden…

NOT EVEN THE PLAYSTATION 5 IS SAFE 🚨 Someone jailbroke a PS5 and got Qwen 3.8 27B running on its GPU at 20 tok/s. That's a real 27B AI model running locally on a games console. The same model on the PS5's CPU did about one token a minute. Meanwhile, DDR4 prices are up 24% in 3 months, AMD just raised GPU memory prices, and 5090s are basically gone. PS5 coming to a datacenter near you.

【西川和久の不定期コラム】9年前の「Tesla V100」がアツい。Qwen3.8-Flash-Nextが約80tok/s出る！ pc.watch.impress.co.jp/docs/…

My home AI lab: two DGX Station GB300s, each with an RTX PRO 6000 sidecar, three DGX Sparks, and a Threadripper linux server, Synology NAS and Mikrotik, all on solar and batteries. Local for everything you can, frontier where it still wins. What it's done so far: -DeepSeek-V4.1-Flash: 5,774 tok/s at 64 users across two Stations, about 430 tok/s single-user, GSM8K 98-99% -DS V4.1 Flash: One Station with a 6000 sidecar: 1,856 tok/s at 16 users, and 2.7-3.25x faster at 16-32 users than without the sidecar -DeepSeek-V4-Pro (1.6T) on two Stations: 83-95 tok/s single-user, 274-313 at 16 users -MiMo-V2.6-Pro: 503 tok/s at 16 users, 1.81x the same pair without sidecars -Nemotron Labs 3 Coding 550B: 99 tok/s single-user, 889 at 32 users -Nemotron 3 550B on one Station, stock vLLM: 40 tok/s single-user, 32 users served -GLM-5.3-Flash on one GB300 with @ashxhart's TensorFold + @MiaAI_lab's recipe: 181 tok/s single-user; full GLM-5.3 across two Stations: 73 -Speculative decoding measured honestly: 1.77x on real text, slower on random tokens Every number is in the public repo: github.com/RonanLabsAI/dgx-s… I run a DGX Station owners group. DM me if you own one.