Founder, @WynfallAI. Building our own AI models in public. Your work stays yours. Opening to everyone soon.
Georgia, US
Joined May 2013
- Tweets2.2K
- Following391
- Followers1.5K
- Likes3.5K
Same two RTX 3090s. Same 280W per card. Same 27B INT4 model.
One change to how the GPU comes up, applied before the driver loads. No extra watts!
BF16 GEMM: 63.7 to 93.1 TFLOPS per GPU (+46%)
TF32: 34.9 to 54.8 (+57%)
FP16: 62.3 to 88.3 (+42%)
Prefill, fresh prompt, no cache:
256K: 666 to 829 tok/s (+24%)
63.5K: 913 to 1,094 tok/s (+20%)
7.4K: 1,025 to 1,198 tok/s (+17%)
Decode:
256K: 95.4 to 104.8 tok/s (+10%)
63.5K: 121.9 to 126.7 tok/s (+4%)
7.4K: 132.7 to 131.7 tok/s (flat, inside the sample range)
First token on a cached 256K prompt: 4.38s to 3.62s (-17%)
Repo: github.com/Zagitta/gpu-unloc…
Credit: @superalesha
If you run 3090s, what are you seeing?
Scaling compute in public is brutal and anyone building in AI knows the pain. But for people relying on these models for actual daily operations, predictability is everything.
If infra limits force a drop from 20X to 10X, leading with the capacity bottleneck builds way more trust than spin. Temporary credits are nice but stable baselines are what we actually build on.
I’ll explain the new Pro 200 plan differently, before I start live tweeting from DevDay on things that are going out!
Today we are going to ship a number of things that increase what you can do across the Plus and Pro plans. A lot of compute is online for this increase. As we increase the floor, we are changing the relative difference between plans to be
Plus = 1X
Pro 100 = 5X
Pro 200 = 10X
and we are reopening subscriptions for Pro 200 (we had paused it). If you have an existing plan you will keep the 20X multiplier for a bit and also receive a lot of additional credits because we know changes are hard even if it means that everyone will get more in the end.
DEEPSEEK V4.1 FLASH ON 2 x DGX SPARKS
This is why DGX spark prices are going to go up and up same with other GB10 but this is just the start of the market before it gets too crazy🤯🤯
Fresh signed llm-speed run: gemma-4-12b Q4_0 on RTX 5090 hits 141.1 tok/s in llama.cpp. Chat feels native.
Public card from today, not my bench. Same suite also shows ~60 ms TTFT on the short chat workload.
llm-speed.com/r/r_v5arqmazf3…
Raising timeouts from 2h to 10h changes what fail means. 82.6% on Coding Intelligence Index v4 is more honest if older runs were just timing out mid-task.
Okay, since my Fable 5.1 benchmark is going to take over 100 hours to run, I thought I'd get some results out as it's going.
And I'm pretty excited about this since it's the first full scale test of my new Coding Intelligence Index v4 at @VulcanBench
This run is Fable 5.1 at High effort, and it scores an 82.6%, with all four failures being real failures, not timeouts.
With this new eval suite I have increased timeouts from 2 hours to 10 hours to make sure if a task fails, it because it failed to solve.
This test was done in the Claude Code harness. Given how expensive it is to run this model, I will not be doing a harness comparison unless I win the lottery 😅
Low, Medium, Extra, and Max still coming.
Once the full effort sweep is in all results will be on the VulcanBench site, and the Coding Intelligence Index v4 leaderboard build process will begin.
Live long and benchmark 🖖
Qwen3.8-Max just took the WebDev Arena lead. The useful part is not the rank, it is that coding evals are now moving faster than the model-size charts people still use as shorthand.
🚀Qwen3.8-Max just got upgraded. Meet Qwen3.8-Max-0902!
2.4T parameters. 1M context tokens. Built for real world complexity.
Further post trained on Coding & Cowork, Qwen3.8-Max-0902 now delivers stronger performance across complex enterprise tasks, scientific research, and long horizon workflows.
💰Pricing per 1M tokens:
$2 input, $6 output.
$0.17 explicit cache hit, $0.25 implicit cache hit.
Now live via API on QwenCloud. Come try it! 🙌
API: qwencloud.com/models/qwen3.8…
Apple Silicon inference got a real open alternative today. Perplexity's Lily chart shows 1.23x higher prefill and 1.35x higher decode than MLX-LM on the same Qwen3.6-35B-A3B setup. The runtime matters as much as the hardware.
Pretty good from Muse Spark is doing a lot of work in that video.
Curious where it falls apart vs Fable 5.1 once the task isn't a clean demo.
Your local model wasn't slow. Your SSD was.
That giant lookup table for Qwen Flash-Next was getting read in tiny scraps. Fix the disk path and Spark jumps from ~300 to ~750–800 tokens/sec on the prompt. Same box. Same model.
🤯 A llama.cpp tweak made Qwen3.8-Flash-Next 2–3× faster at prompt processing on a DGX Spark.
And the optimization is basically is to optimize SSD reading. Seems almost trivial 😂
Since the Qwen3.8-Flash-Next has a HUGE PLE/n-gram table that can live on SSD instead of consuming precious RAM/VRAM.
However, llama.cpp's current mmap path can trigger thousands of tiny scattered page faults, often pulling 4KB when the model may only need ~90 bytes.
PR #28136 introduces a new direct-read path using parallel pread() workers.
Reported DGX Spark results:
🐢 Before → ~300 tok/s prefill
🚀 After → ~750–800 tok/s
Cold diverse-text tests show roughly:
2.0–3.1× faster prompt processing
WITHOUT loading the giant PLE table into RAM.
🧠 Strix Halo 128GB testers are reporting +20–32% cold prefill too.
⚠️ PR #28136 is currently OPEN and not merged yet.
🔗 llama.cpp PR #28136 👇
Corey Higgins retweeted
🎬 NextLVL is Hiring Video Editors! 🎬
We’re looking for talented and reliable video editors to join the team. This is a long-term role with weekly deliverables for top Fortnite creators.
What We’re Looking For:
• Joins all team meetings (4x/week)
• Attends 1-on-1 check-ins (1x/week)
• Works ahead on assigned videos
• Self-sufficient and proactive
• Strong team player
• Gathers fresh, in-game B-roll for each script
• Delivers high-quality short-form edits consistently
If you're passionate about Fortnite content and want to grow with a team that’s pushing the scene forward—DM me or tag someone who should apply.
Let’s level up together 💥
Example video:
Left @2AM_Esports on great terms. Looking forward to the future of getting back into content ❤️
Corey Higgins retweeted
WALLHACK Gravity Yume Giveaway
1x SP-004 Gravity Yume Mousepad
Steps to enter:
🔹Follow @WALLHACKHQ
🔹Like and RT the post
🔹Tag one friend
🏆One Winner
🗓️Ends April 22nd
Big things coming! already backed by top brands. Looking to partner with more sponsors for future content/events. If you want in, DM me. Let’s create something epic together!
To all Creative’s Out There! 🎨
Looking for a skilled graphics/thumbnail designer to take my future Youtube videos to the next level!
📩Looking forward to DMs with portfolios,rates, and other work samples. Can't wait to join creative forces and build something amazing together!
If some of you don't know I started making music. Check it out in Killa's montage!
Preview For @FNCompetitive
Beneath it All ❤️🩹 ft @zikoxo
@2AM_Esports
Full 4K Video ❤️ + ♻️
youtu.be/wUBipGzvHhc?si=O24U…
