🎮 Tech, gaming, AI, and everything in between. 🤖 Building with it, not just talking about it. 🔥 From the mind of @ToNYD2WiLD
Joined March 2026
- Tweets4.1K
- Following222
- Followers4.8K
- Likes1.1K
Tech2Wild retweeted
Big Announcement: Meet Imagination Studio🎬🎵
by @u1tra_instinct and Me!
Our local open source all in one creative platform on your own hardware has been fully redesigned and rebuilt from the ground up. Video,Audio, Music, Image. Releasing ASAP with latest+best local models.
Tech2Wild retweeted
I've built 3 channels in 3 different niches: sneakers, guns and tech. These 5 rules worked every time 👇
What's your niche?
Tech2Wild retweeted
A friendly reminder to our ever-vigilant watchdog, Mia: Git has a timeline, not just screenshots. How thoughtful of you to remind everyone to check yours, too.
Funny how this brings back memories of yesterday.
Anyway, back to work.
This quoted post is unavailable.
Tech2Wild retweeted
Credit where it's due. Others mapped this ground first:
- @Tech2Wild documented the same clock latch: github.com/tonyd2wild/DeepSe…
- @z841973620 decoded the EC's automatic fan curve in dgx-spark-fan-override: github.com/Z841973620/dgx-sp…
- antheas wrote spark_hwmon, the power driver mine builds on: github.com/antheas/spark_hwm…
Tech2Wild retweeted
Normalize calling people out when they’re doing shady stuff (stealing others work and efforts to name a few)
Then this person had the audacity to say “Do better, Ahmad” when I called them out on it
Opensource runs on trust PERIOD
They built a 37K cult using a fake avatar to hijack the algorithm. Code never lies and who is actually behind it is revealed inside.
Model credit: @ViC305 & @Blackfrost_ai (not participants).
Read the receipts behind Mia’s AI Lab 👇
Tech2Wild retweeted
The appropriate way to respond to this ⬇️
It's understandable when your agent pulls in code and maybe you didn't give it the proper instruction to credit.
It's NOT understandable when you realize and instead of fixing, you (or your "fans") justify and rationalize the "mistake" that was made <3
In the second situation, there is no mistake.
HUGE respect for @Tech2Wild here 👑
Addressing this once.
If I ever miss a credit on something of yours, DM me. I'll fix it same day. No need to quote tweet.
What happened was I pointed Claude at @majewskizby GLM stack, folded it into my recipe and ported it from 4 Sparks to 2. knapcio is credited in the tweet, at the top of the repo as the new default, and in the Credits. He credits my image in his repo too. Credit goes both ways.
His repo credits the people he built on, Matt included. So the real question for the community, if I use someone's recipe, do I also credit everyone THEY credited, and if so do I need to go credit the person they credited their creditors too ? That's a family tree with no end.
I'm not the wizard and never claimed to be. None of this is mine or yours, it's open source. I'm learning in public and building on the backs of this community, proudly. This is LAME
Tech2Wild retweeted
Actually, what’s disappointing is people like you who all they care about is getting credited and decided to publicly feud about it online. We are all standing on the shoulders of giants so drop your narcissism and ego, it is not welcome here.
Tech2Wild retweeted
People will HATE A.I. and say it is taking jobs and they rather pay a human to do the work. Not knowing the human they are paying to do the work is using A.I.
Tech2Wild retweeted
MiMo-V2.6-Flash on 2× DGX Spark, day-zero,
@Tech2Wild vLLM recipe. 44 tok/s
Rocket-launch brief, thinking on, no cap: 80.3K tokens in 30 min: 0.5 s TTFT, DFlash accepting 4–5 tokens/step in the code phase. Complete file, first try, zero errors.
Same prompt on GLM-5.3-Flash, same pair: 79.6K tokens, 57 min, at about 23 tok/s.
Same thinking budget, half the wait. MiMo is way faster but GLM's video is a bit better:
Richer pad, more dramatic climb. Top GLM, bottom MiMo.
🔥 MiMo-V2.6-Flash-RL 310B · 2x DGX Spark TP2 · vLLM + DFlash (Checkpoint)
Benched warmed + streaming, fixed 9-category prompt set, C1 to C6:
⚡️ 70 code · 69 math · 57 JSON · 26 prose · 88 tok/s single-stream on counting · 85 structured tables ·
🎯 DFlash k=7 runs hot: up to 6.9 of 7 draft tokens accepted per step on predictable text
🚀 156 tok/s aggregate at 6 streams · 206 on code · 249 on tables
🧠 1.87M-token fp8 KV · 300K ctx serving · 6 full 300K requests at once · 0.37s TTFT
github.com/tonyd2wild/MiMo-V…
Tech2Wild retweeted
I always seem to have more success with @Tech2Wild recipes. Reliable and well put together and always very quick support. Give him a follow if you haven’t already for your DGX sparks and 3090 goodies. 🔥
UPDATE !
🔥 DeepSeek-V4.1-Flash · 552B MoE (769B w/ Engram) · MXFP4/MXFP8 · 4x DGX Spark TP4 · vLLM + DSpark
Day-0 bring-up, now on boot 10 🚀
⚡️ 73.8 tok/s CODE, one stream (+41% vs last boot)
🔢 92 tok/s counting · 55 tables · 52 JSON · 51 math
📈 225 tok/s code at 6 streams · 260 peak aggregate · 132 avg across 8 categories
🧠 1.07M-token KV · 300K ctx · 1M proven · tools + vision on
⏱️ 0.3-0.5s TTFT · prefill up to 1,539 tok/s (93K-token prompt in 78s)
💾 203GB of Engram tables served from disk, now node-local on every Spark
github.com/tonyd2wild/DeepSe…
Tech2Wild retweeted
Wow intraday improvement and double the speed 🫡 @Tech2Wild if you not following him and you have dgx-sparks you may be missing out some the lastest and greater that you can deploy on your sparks cluster
UPDATE !
🔥 DeepSeek-V4.1-Flash · 552B MoE (769B w/ Engram) · MXFP4/MXFP8 · 4x DGX Spark TP4 · vLLM + DSpark
Day-0 bring-up, now on boot 10 🚀
⚡️ 73.8 tok/s CODE, one stream (+41% vs last boot)
🔢 92 tok/s counting · 55 tables · 52 JSON · 51 math
📈 225 tok/s code at 6 streams · 260 peak aggregate · 132 avg across 8 categories
🧠 1.07M-token KV · 300K ctx · 1M proven · tools + vision on
⏱️ 0.3-0.5s TTFT · prefill up to 1,539 tok/s (93K-token prompt in 78s)
💾 203GB of Engram tables served from disk, now node-local on every Spark
github.com/tonyd2wild/DeepSe…
Tech2Wild retweeted
4 DGX Sparks running DeepSeek V4.1 Flash
6 concurrent sessions, 170.73 tok/s coding throughput. 🔥
1,200 output tokens in 7.028 seconds.
Warm single-stream:
💻 Code: 77.61 tok/s
🧮 Math: 74.08
🧠 Reasoning: 48.75
✍️ Prose: 33.02
6-stream aggregate:
💻 Code: 170.73 tok/s
🔢 Counting: 291.64 tok/s
552B MoE · MXFP4 · TP4 · vLLM + DSpark + CUDA graphs
✅ Tool calling + vision
✅ 127,480-token retrieval test passed
✅ 300K configured context
Video replays the actual recorded streams at 1× speed. Thinking OFF, 200-token cap per coding request.
Short-prompt coding throughput
Credit to @Tech2Wild for the Boot9 recipe.
@NVIDIAAI
Tech2Wild retweeted
NVIDIA GLM-5.3 Flash NVFP4 on 2x DGX Spark, TP2 over RoCE.
DFlash2 took median decode from 15.18 to 20.50 tok/s (+35%). Passed retrieval across 120,124 input tokens.
Our work: TP2 memory/startup fixes, depth sweeps and raw-stream verification. Four sustained requests per arm; all stopped naturally.
Built on @Tech2Wild GLM vLLM image + SM121 sparse-indexer patch.
Recipe, raw streams and failure records:
github.com/Weschera/GLM-5.3-…
I know there a lot of recipes out there but i wanted to share one with the official @NVIDIAAI quant. Still a lot of room for improvement
Tech2Wild retweeted
My friend @Tech2Wild came through with a huge gift. Saving everyone so much time to have a working fast recipe for TP4. LFG. Time to start ablit
💪
🔥 DeepSeek-V4.1-Flash 552B MoE (769B w/ Engram) · MXFP4 · 4x DGX Spark TP4 · vLLM + DSpark (Checkpoint)
Benched warmed + streaming, fixed 8-category prompt set, C1 to C6:
⚡️ 40.5 tok/s single-stream on counting · 39 structured tables · 33 code · 31 math · 15 prose
🎯 DSpark k=5 runs hot: up to 5.6 of 6 draft tokens accepted per step on predictable text
🚀 190 tok/s aggregate at 6 streams · 138 on code
🧠 841K-token KV · 300K ctx serving · 1M ctx proven (1.08M-token pool) · 0.44s TTFT
💾 The 203 GB Engram tables stay on disk, so 552B fits in 4x 128 GB unified memory
🛠️ 5 tok/s eager ➜ 41 tok/s with CUDA graphs + DSpark, after 8 fixes: Engram-on-disk + rank-offset bug, graphs around a host-side lookup, GB10 top-k + indexer page fixes, block size, JIT wedge
Day-0. 552B-class reasoning on desk hardware 🖥️
github.com/tonyd2wild/DeepSe…
Tech2Wild retweeted
Since X Money is limited to one account per person and is now required to receive payouts, U.S.-based creators are therefore limited to one monetized X account.
Say goodbye to the days when people could run multiple, monetized accounts.
Tech2Wild retweeted
we're now moving into the AGI era (whether you view it as this model, the last one, or the next one), and could not do it without close partners
GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years.
AGI has arrived. Congratulations @OpenAI team.
400K GPUs coming online next.
Tech2Wild retweeted
@Tech2Wild Kudos on this recipe! Hands-on testing on my DGX shows it handles Thai super smoothly, and output quality is by far the best I've seen so far. The new NVIDIA NVFP4 is likely doing some serious heavy lifting here.
