Here is my first article which talks about how I am running DeepSeek on two of my DGX Sparks:
gattic.com/articles/two-dgx-…
More articles to come! If you have any topics you would like me to write about, let me know!
I am really excited about Hy4. I am just surprised at its size...
⚖️ Tencent gave the Local AI crowd a 770B to figure out how to run!
😬And the smallest published GGUF is only ~215 GiB.
Can we squeeze it onto a DGX Spark?
The weights are finally here!
Tencent (@TencentHunyuan) released its flagship Hy4 open weights, and AngelSlim already has 3x GGUF quantizations available.
Here are the model stats as a refresher
🧠 770B total parameters
⚡ 49B active per text token
📚 1 MILLION token context
👁️ Native multimodal architecture
🖥️ GUI and computer use
🎮 Unreal, Unity, Godot, Blender and CAD workflows
🚀 Built-in 3.3B DFly2 speculative drafter
GGUF sizes!
💾 Q4_K_M → 437.93 GiB
💾 IQ2 → 309.58 GiB
🔥 STQ1_0 → 214.65 GiB 👈 that smallest version averages just 2.39 bits per weight!
🍎 Mac Studio M5 Ultra, 512GB unified RAM → Enough capacity for the Q4 weights on paper.
🍎 256GB Mac Studio → Potentially enough memory for the aggressive STQ1_0 version.
🟢 4× DGX Spark, 128GB each → 512GB combined memory, but distributed inference support is required.
And here's an interesting economic comparison.
512GB Mac Studio → $22,299
4× DGX Spark at $6,950 → $27,800
💰 Difference → $5,501
The STQ1_0 quant also requires a custom llama.cpp patch, and its coding/reasoning quality needs testing.
⚠️ Licensed under Tencent Hy License, not plain Apache 2.0.
Links in ALT.
5
59
Robert retweeted
Short highlights video of my interview with @coinbase @base @baseapac is live on my YT channel
Was a pleasure to sit down with @Nibel_eth and talk about how my biggest L ended up as the biggest push in my crypto journey. It’s all about the execution. Trench smart
Link below 👇
Finally Sol and Astra 6.2 are available!
Some of you have been asking very politely. Some of you have just been asking.
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
openai.com/index/introducing…
1
46
Robert retweeted
A new open-weights DeepSeek V4.1-Flash competitor just appeared.
It fits on 2x DGX-Sparks and can ...
operate scientific software, check its own work, and FIX its mistakes.
It uses the same architecture as GLM-5.3-Flash.
This is Sage-S2.1-Flash from Shanghai AI Lab.
Here are the model's stats ...
🧠 320B total / 18B active
📚 ~1M token context
👁️ Multimodal
🔓 MIT license
💾 BF16 + FP8 weights downloadable
Shanghai AI Lab demonstrated it working with 12,487 mixed medical images.
Example application, the model
👉 Identified 139 relevant CT slices
👉 Attempted a reconstruction in 3D Slicer
👉 Detected a slice-ordering inconsistency
👉 Corrected the error
👉 Reconstructed the skeleton in 3D
It didn't just describe the images.
Also, it operated scientific software and used verification feedback to correct its work.
And Sage supports 10+ types of scientific inputs
🧬 DNA / RNA / protein sequences
🧪 Molecular graphs
💎 Crystal structures
📈 Spectra
🧠 EEG / ECG signals
Now look at these developer-reported benchies.
MLE-bench Lite:
🔥 Sage S2.1 Flash → 90.3
🐋 DeepSeek V4.1 Flash → 89.2
🟠 Claude Opus 5 → 88.0
⚫ GLM-5.3-Flash → 84.3
🟣 Qwen3.8-Max → 82.2
🟢 GPT-5.6 Sol → 80.7
And compared with GLM-5.3-Flash:
OfficeQA Pro
64.7 → 72.2
Harvey LAB
20.0 → 25.0
SpreadsheetBench 2
32.4 → 37.1
Terminal-Bench 2.1
84.3 → 87.9
Not a clean sweep across every benchmark, but these are interesting results for a scientific agent.
⚖️The weights are already here.
💾 BF16 → ~645GB
💾 Official FP8 → ~330GB
Raw weight math:
4-bit → ~160GB
3-bit → ~120GB
2-bit → ~80GB
Remember all the work Localmaxxers have done getting GLM-5.3-Flash running on affordable hardware?
So who's going to quantize it for a 192GB Gorgon Halo?
Or two DGX Sparks? 👀
Imagine running your own scientific research agent that doesn't just generate answers.
It executes experiments, checks results, and corrects its mistakes.
Link in reply










