@Dataphil3

Building sovereign SI infrastructure. 15-node cluster · 6× RTX PRO 6000 Blackwell (576GB VRAM) · 2TB ECC RAM · PB-scale Ceph · 200GbE fabric

 Mountain View, CA / London, UK
puretensor.si
Joined November 2013
  • Tweets232
  • Following1.5K
  • Followers1.4K
42Media
Been using this for a few hours , great first impression! @AnthropicAI really cooks ! first thing I noticed, it's actually a duet. you and claude. no swarm feeling, no clogging up everything with @grok bots or dots, lots of dots. if theres subagents they're backstage, that’s the first impression
Claude
@claudeai22h
We’re opening a research preview of Claude Duet, a new app where Claude works as a proactive agent. It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf.
3
1
3
283
Not the kernel. Our PCIe. @fraserpricee's fix, one reboot: IOMMU off + NVIDIA P2P driver keys. 237 -> 444 tok/s, his exact flags 461 tok/s, best config 414 verified on LocalMaxxing Frontier-class @deepseek_ai on one workstation. Thanks Fraser. @NVIDIAAI #DeepSeek #LocalLLM
7
5
1
61
9,628
The box: 4x RTX PRO 6000 Blackwell Max-Q (96 GB, 325 W, PCIe 5.0 x16), Threadripper PRO 9975WX on ASUS Pro WS WRX90E-SAGE SE, 512 GB Samsung DDR5 ECC, 3x Samsung 9100 PRO Gen5 NVMe RAID0 for weights and Engram. Verified runs: localmaxxing.com/runs/cmvins…

DeepSeek-V4.1-Flash-1017 — 413.6 tok/s on RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB...

DeepSeek-V4.1-Flash-1017 achieved 413.6 tok/s output speed on RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB ×4 using vllm · FP8. See the full speed test on LocalMaxxing.

localmaxxing.com
1
7
1,044
Same cards. Same image. Same flags. Different kernel. Reran @fraserpricee's @deepseek_ai V4.1-Flash-1017 setup: 240 tok/s here, ~600 his 5.35 vs 5.25 tokens/step 21 ms vs 8.8 ms per step His kernel is the secret sauce. Waiting on that recipe. @NVIDIAAI #DeepSeek #LocalLLM
4
1
5
443
The math, batch 1, temp 0, his exact prompt: tok/s = tokens per step / step time 5.35 / 22.3 ms = 240 (us) 5.25 / 8.8 ms = ~600 (him) Same acceptance, same prefill. The whole gap is time per step. Our runs (3 verified): localmaxxing.com/runs/cmvikw…

DeepSeek-V4.1-Flash-1017 — 239.6 tok/s on RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB...

DeepSeek-V4.1-Flash-1017 achieved 239.6 tok/s output speed on RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB ×4 using vllm · FP8. See the full speed test on LocalMaxxing.

localmaxxing.com
83
Dataphile retweeted
NVIDIA AI
@NVIDIAAI23h
Replying to @mervenoyann @llm_wizard
let's goooooo
1
2
11
682
📣 @googlecloud announces general availability of NVIDIA Vera Rubin infrastructure. Coming next month, A5X instances bring together NVIDIA Vera Rubin NVL72, NVIDIA ConnectX-9 SuperNICs and the Google Cloud Virgo Network for training, inference and large-scale reinforcement learning. Discover how @IneffableLabs is advancing “superlearner” systems and how @WPP is bringing next-gen creative agents to A5X using NVIDIA Nemotron and NVIDIA OpenShell. ➡️ nvda.ws/3VYBtRJ
5
18
202
4,270
Gemini 4 Argon built a 3D periodic table synced to Tom Lehrer’s song, “The Elements.” Created in @Antigravity, it animates orbiting electrons, atomic data, and light spectra.
314
549
139
4,600
547,663
Dataphile retweeted
NVIDIA
@nvidiaOct 21
49
141
29
1,624
66,956
Dataphile retweeted
NVIDIA GTC
@NVIDIAGTCOct 21
The #NVIDIAGTC Berlin keynote is live. 🎤 Watch NVIDIA CEO @JensenHuang at Tempodrom unveil the latest breakthroughs in AI infrastructure, agentic AI, physical AI, and more. Tune in now: nvda.ws/4yvMNSU
8
38
3
327
18,961
Dataphile retweeted
NVIDIA GTC
@NVIDIAGTCOct 21
The #NVIDIAGTC Berlin Pregame has just begun. 🎬 Industry leaders are on the live stage setting the conversation before Jensen Huang’s keynote. Tune in now → nvda.ws/4yvMNSU
4
3
2
43
2,403
Dataphile retweeted
NVIDIA
@nvidiaOct 21
With NVIDIA extreme co-design, we’ve accelerated AI performance a millionfold over 10 years, and we’re not stopping there.
102
162
25
1,865
96,059
Dataphile retweeted
NVIDIA
@nvidiaOct 21
Doing what was previously unimaginable 🚀
66
225
31
3,383
153,049
#888 a lucky number
2
1
17
411
Half a terabyte of DeepSeek. One box. @deepseek_ai V4.1-Flash-1017 on 4x RTX PRO 6000 Max-Q: 463 tok/s aggregate 102 tok/s per stream 93/95 real bugs fixed In the other corner: @nvidia Nemotron-3.5-Super-VL on 2 cards, 1,535 tok/s. puretensor.ai/blog/what-fits… #DeepSeek #LocalLLM
7
4
1
27
1,274
Six suites, three built from our own data. DeepSeek-V4.1-Flash-1017: 97.9% fixing 95 real bugs 95.0% on AIME 2026 92.3% reasoning 91.1% long context, 32k to 222k Nemotron-3.5-Super-VL led vision (94.8% vs 90.5%) and hit 93.3% on AIME 2026, from 2 GPUs at 15% of the size.
92
You can stack the #RTX6000PRO workstation edition tightly tightly together see here below , as long as you set the mas Wattage to 250W. This way the thermals stay within a reasonable range. @NVIDIAAIInfra @AMD @Samsung #ConnectX7
6
2
15
850
@nvidia #ai
1
30
@Noctua_at
3
169
Dataphile retweeted
Anthropic
@AnthropicAIOct 13
New on our Engineering blog: Creating in-house source control. In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale. Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: anthropic.com/engineering/re…

Redesigning source control at Anthropic

As our merges per day grew more than tenfold, source control became our biggest development bottleneck. Here, we share how we migrated to an in-house version control system in five months.

anthropic.com
1,174
2,663
2,193
25,990
8,285,734
Load more