
Dad and Husband. Local AI - DGX Spark + 48GB M5 Pro. TensorFold contributor. Building @MuraTalkApp and some other things. Ngunnawal land 🌏 Views my own.
Canberra, Australia
Joined October 2026
- Tweets1.8K
- Following442
- Followers407
- Likes2K
Ripple-65B Preview is a big dense model. No experts to skip, so every token reads the whole thing.
80 layers, 12T training tokens, 1M context, and 80.5 on Terminal-Bench 2.1 (publisher's score).
In my evaluations it outscored Qwen3.8 27B and Flash-Next, 58/75 vs 48 and 47, with the biggest gap on the coding tasks: 8/15 vs 0/15 for both Qwen models.
The NVFP4 release is conservatively quantised. The MLPs drop to 4 bits but attention stays at 16, so it's 57 GB and fits on one DGX Spark with room to run.
Against the publisher's model code run in float64, it gets 98.7% top-1 agreement and a mean KL of 0.0007.
At the Spark's memory bandwidth of 273 GB/s, reading roughly 56 GB a token caps you near 4.9 tok/s. I measured 4.1.
Not so long ago, that was the end of the story. But there's more to it 👀
Humanity is sick.
Does he have the cure?
Take a thought-provoking and compelling journey through the limits of wellness, longevity and alternative thinking.
Podman, a new documentary from independent Australian filmmaker (and my childhood friend) Simon Dall.
Out now.
youtu.be/2updBsuit0k
1
61
Full documentary available for free on YouTube.
youtu.be/o6547b_8e-w
34
Wisdom from Elshayib here.
Knowing how AI works - and understanding at a fundamental and technical level both the benefits and the risks - is going to make you incredibly important in the medium to long term.
We're so lucky to be at the heart of the AI industry, informed and updated about the most important technology of our time.
The technology that's reshaping every aspect of our lives and potentially the most beneficial and yet dangerous to ever exist.
A lot of you are more important than you think as we play a big role in all of this whether we use AI or build it or even simply talking about it.
2
2
87
Now this is interesting!
228bn parameter flash model. 68bn of them on n-gram. Small enough to fit on a single DGX spark with SSD offloading of the n-gram table.
Matches Haiku 5.5 on high effort, with open weights coming soon.
The Chinese labs are lifting their game big time 🚀🚀
⚡Meet Hy4 flash, the cheapest model for everyday productivity. Matches Claude Haiku 5.5 (high) on daily work (human blind tests), yet 18x task cost reduction. API pricing (per 1M tokens): $0.003 cached · $0.070 input · $0.281 output. 🔗:hy.tencent.ai/research/hy4-f…
Bit of a self-indulgent one, sorry 😅 but, with a heavy dose of gratitude, I wanted to take a moment to look back on the week that was.
283 to 403 followers. 384 replies to 199 of you. over 90,000 views, and 77% of them came from just chatting in the replies.
Biggest thanks to the people I talked with most: @carthurj @MiaAI_lab @ashxhart @CerebralCoding_ @TekonBurns @TheAmazeX @NeoAIForecast 🙏
A month ago this account was zero. Grateful for every one of you. Back to building 🛠️
4
6
210
It’s here! Qwen 4 Max has been released!
It OUTPERFORMS opus 5.5 on several key benchmarks.
And yes, Qwen says open weights of smaller models are coming 🚀🚀🚀👀
Qwen 4 27B is near!
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
3
143
Good point from @TekonBurns. Metering by the token means the provider has an economic incentive to maximise token use.
With local models your incentive is to run the most efficient system possible.
List price per million tokens is officially a trap.
The New Stack ran Claude Fable 5.2 against GPT-6 Astra on three real coding and billing tasks today.
Fable 5.2 is priced 20% lower per token on paper, yet it burned 5x more tokens, ran 2.6x longer, and billed 236% higher at $13.39 versus $3.99.
Token-efficient reasoning models mogging verbose thinkers.
1
1
198
*Insert 404, followers not found joke here* 😂
Thank you to all 404 beautiful humans who've followed along. I'm loving the connections and having a space to geek out and push what's possible with local AI.
Just over a month ago that number was zero. It only went up because you gave me a chance.
Thanks for being here 🙏 And if you like what I'm serving up, drop me a follow!
6
262
Omg. Just watch it. How is this even real.
Hey @grok, on a scale of 1 to 100, with 100 being the hardest, how difficult was this Starship catch that SpaceX pulled off today?
1
2
105
By popular demand, the ultimate showdown: M5 Pro vs DGX Spark 🥊
Qwen3.8-27B, same TensorFold 1.2.8 harness. At 1 stream, Mac writes a bit faster, 111.8 tok/s vs 95.7 on code. At 8 streams, Spark leads 328.3 to 205.2, or 1.60x. Spark reads 8k 2.9x faster and 120k 4.0x.
Spark has other perks too, like 128 GB unified memory and a 4 TB NVMe SSD. But on a like-for-like basis, compute is where you clearly see the gap. Bandwidth is similar (307 vs 273 GB/s). MLX 4-bit on Mac, NVFP4 on Spark.















