Dataphile (@Dataphil3) header
Dataphile (@Dataphil3) profile picture

Dataphile

@Dataphil3

Building sovereign AI infrastructure. 20-node cluster · 6× RTX PRO 6000 Blackwell (576GB VRAM) · 2TB ECC RAM · PB-scale Ceph · 400GbE fabric

 Santa Clara, London
puretensor.si
Joined November 2013
  • Tweets601
  • Following3.1K
  • Followers3.1K
68Media
Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)Photo by Dataphile (@Dataphil3)
DeepSeek V4.1 Flash, running on our own GPUs, fills in an expense claim on a Mac and submits it. 26 seconds, real time. cu-local runs @OpenAI's Codex Computer Use with your own model. Open source, MIT. github.com/puretensor/cu-loc… @deepseek_ai #ComputerUse #OpenSource
6
3
19
638
Quantum Computing Became Checkable

Quantum Computing Became Checkable

Article · @Dataphil3
1
6
115
AlphaEvolve and AI That Improves Its Own Stack

AlphaEvolve and AI That Improves Its Own Stack

Article · @Dataphil3
3
3
12
447
The Quiet Rebuild of Financial Plumbing

The Quiet Rebuild of Financial Plumbing

Article · @Dataphil3
1
14
374
Replying to @Alibaba_Qwen
@Alibaba_Qwen 🤨when!
Image from Dataphile (@Dataphil3)
2
1
19
340
Three Countries, Three Ways to Build a National AI Stack

Three Countries, Three Ways to Build a National AI Stack

Article · @Dataphil3
1
2
14
423
FINALLY got into the Original gang & @X is finally starting to pay. #nice
Image from Dataphile (@Dataphil3)
5
1
21
388
The Engineers Turning CO2 Into Products

The Engineers Turning CO2 Into Products

Article · @Dataphil3
2
14
332
For reasons I won’t go into . I have so much frontier intelligence credit right now , that I feel like a total Token Glutton , a bloated tick .
3
13
285
Using Opus 5.5 from @AnthropicAI on Fast as orchestrator and GPT-6.1 Sol from @OpenAI on Ultrafast as grunt coding workers feels especially degenerate and decadent.
Image from Dataphile (@Dataphil3)
3
20
416
AI Has Outgrown the Word Unicorn

AI Has Outgrown the Word Unicorn

Article · @Dataphil3
2
3
11
367
Ripple-65B from @RicursiveAI is a good model. We tested it ourselves on one RTX PRO 6000. After warmup, in our agent: steady 50-70 tok/s on prose ("write an essay and a poem about regret") and 110-120 on code ("write a simple python script 1000 lines"). #LocalLLM #OpenSourceAI
Image from Dataphile (@Dataphil3)
2
4
23
1,045
Run pages, Verified on LocalMaxxing (batch 1, temp 0, median of 5): 135.2 tok/s code: localmaxxing.com/runs/cmw2sf… 97.9 reasoning: localmaxxing.com/runs/cmw2sf… 24.0 without the drafter: localmaxxing.com/runs/cmw2s1… The fastest Ripple runs on the site so far, any GPU.
Ripple-65B-Preview-NVFP4 — 135.2 tok/s on RTX PRO 6000 Blackwell Workstation Edition · 96 GB

Ripple-65B-Preview-NVFP4 — 135.2 tok/s on RTX PRO 6000 Blackwell Workstation Edition · 96 GB

Ripple-65B-Preview-NVFP4 achieved 135.2 tok/s output speed on RTX PRO 6000 Blackwell Workstation Edition · 96 GB using sglang · NVFP4. See the full speed test on LocalMaxxing.

localmaxxing.com
1
1
90
The Hidden Supply Chain Behind Every Advanced Chip

The Hidden Supply Chain Behind Every Advanced Chip

Article · @Dataphil3
1
2
9
467
Really impressed with @OpenAI 500$ sub now . I have #Ultrafast on now as default on Sol6.1 xhigh and I am not seeing it eat the weekly limit down so fast as to be unusable. This was the main complaint for ultrafast from multiple users before.
2
14
340
Replying to @Lovable
@Lovable CTO and @steipete at @OpenAI Dev Day in London
Image from Dataphile (@Dataphil3)
Image from Dataphile (@Dataphil3)
2
12
474
I love #ultrafast @OpenAI let’s see how quickly it burns down weekly limit
Image from Dataphile (@Dataphil3)
8
192
The Terminal Became the Most Powerful IDE

The Terminal Became the Most Powerful IDE

Article · @Dataphil3
2
2
8
347
Nemotron-3.5-Lightning on one RTX PRO 6000: 770.4 tok/s on LocalMaxxing's reasoning prompt at batch 1, Verified. The best Verified run on one of these cards before it was 593.3. The change: a 5-token DSpark draft instead of 3, with thinking on. #LocalLLM
Image from Dataphile (@Dataphil3)
5
1
11
348
Run: localmaxxing.com/runs/cmvy5w… Recipe: nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 with NVIDIA's DSpark draft, num_speculative_tokens 5, upstream vLLM nightly. Thinking on (template default), temp 0, median of 5. 593.3 was gpu-rentals' run with K3.
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 — 770.4 tok/s on RTX PRO 6000 Blackwell Workstation...

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 — 770.4 tok/s on RTX PRO 6000 Blackwell Workstation...

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 achieved 770.4 tok/s output speed on RTX PRO 6000 Blackwell Workstation Edition · 96 GB using vllm · NVFP4. See the full speed test on LocalMaxxing.

localmaxxing.com
1
3
114
Anthropic's $80 Billion Loss Is an Accounting Artifact

Anthropic's $80 Billion Loss Is an Accounting Artifact

Article · @Dataphil3
1
1
8
232
Nanbeige4.3-33B on one RTX PRO 6000 Max-Q at 325 W: 94.1 tok/s on code, 92.6 on reasoning, batch 1. Stock FP8 does 42.9. NVFP4 weights plus free n-gram drafts take it to 2.2x. ProCreations' DFlash2 drafter pushes code to 105.3. Fastest Verified run on LocalMaxxing, any GPU.
Image from Dataphile (@Dataphil3)
2
1
9
463
Run pages, all Verified: 94.1 code: localmaxxing.com/runs/cmvyw8… 92.6 reasoning: localmaxxing.com/runs/cmvyw8… 105.3 code, DFlash2: localmaxxing.com/runs/cmvyw8… 42.9 stock FP8: localmaxxing.com/runs/cmvywu… Batch 1, temp 0, thinking off, 1,024 tokens, median of 5.
Nanbeige4.3-33B-NVFP4 — 94.1 tok/s on RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB

Nanbeige4.3-33B-NVFP4 — 94.1 tok/s on RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB

Nanbeige4.3-33B-NVFP4 achieved 94.1 tok/s output speed on RTX PRO 6000 Blackwell Max-Q Workstation Edition · 96 GB using vllm · NVFP4. See the full speed test on LocalMaxxing.

localmaxxing.com
1
3
62
Load more