Wësche's banner

Wësche

@WescheNex1q • 3,172 subscribers

Day time artist and night time AI enthusiast/cybersecurity. Building & benchmarking frontier LLMs on 4x DGX Spark + Mac. Creator of Vesica Studio. Houston

Shorts

Qwen 3.8 27b nvfp4 vs Grok 4.6 Someone tell me how this local model that is 100x smaller and 10x slower made a better animation?! Elon Musk help us solve this, both with their highest thinking mode (xhigh vs max)

Qwen 3.8 27b nvfp4 vs Grok 4.6 Someone tell me how this local model that is 100x smaller and 10x slower made a better animation?! Elon Musk help us solve this, both with their highest thinking mode (xhigh vs max)

166,698 просмотров

Day-0 Qwen3.8-27B vs Qwen3.6-27B: the voxel pagoda test. Same prompt, one attempt each, single DGX Spark, both NVFP4. 3.8: 75,291 tok · 52m 56s · 23.7 tok/s (vLLM, native MTP) 3.6: 12,117 tok · 5m 13s · 38.8 tok/s (llama.cpp, DFlash K10) 3.8 thought for 160K characters before writing a line of code.

Day-0 Qwen3.8-27B vs Qwen3.6-27B: the voxel pagoda test. Same prompt, one attempt each, single DGX Spark, both NVFP4. 3.8: 75,291 tok · 52m 56s · 23.7 tok/s (vLLM, native MTP) 3.6: 12,117 tok · 5m 13s · 38.8 tok/s (llama.cpp, DFlash K10) 3.8 thought for 160K characters before writing a line of code.

102,648 просмотров

Two models. One DGX Spark One identical voxel Eiffel Tower prompt. Qwen3.8-27B xhigh: 83K tokens, 41m29s, 33.4 tok/s Ornith 1.5 35B-A3B: 47K tokens, 10m35s, 74.0 tok/s Thinking on, no token ceiling, matched camera. Ornith finished much faster; Qwen spent far more tokens.

Two models. One DGX Spark One identical voxel Eiffel Tower prompt. Qwen3.8-27B xhigh: 83K tokens, 41m29s, 33.4 tok/s Ornith 1.5 35B-A3B: 47K tokens, 10m35s, 74.0 tok/s Thinking on, no token ceiling, matched camera. Ornith finished much faster; Qwen spent far more tokens.

51,527 просмотров

Local showdown: Qwen vs Deepseek fireworks one-shot: "build a fireworks show over a city, single HTML file, no libraries." two locals on DGX Sparks, temp 0.6, one attempt, no edits. DeepSeek-V4-Flash (2x GB10): 1m20s, 5,214 tokens, Qwen3.8-27B NVFP4 (1x GB10, xhigh): thought for 21 minutes before typing one character of the answer. 38,269 tokens total, ~29K of them thinking

Local showdown: Qwen vs Deepseek fireworks one-shot: "build a fireworks show over a city, single HTML file, no libraries." two locals on DGX Sparks, temp 0.6, one attempt, no edits. DeepSeek-V4-Flash (2x GB10): 1m20s, 5,214 tokens, Qwen3.8-27B NVFP4 (1x GB10, xhigh): thought for 21 minutes before typing one character of the answer. 38,269 tokens total, ~29K of them thinking

40,134 просмотров

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. NVIDIA AI

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. NVIDIA AI

24,282 просмотров

Got Aleph Alpha's new Kolibri 1 running on a single DGX Spark the day it came out. Their container is x86 only, so I built vLLM 0.29.0 (arm64) + their aleph-alpha-inference plugin. FP8 weights (74 GB) and an FP8 KV cache, 164K context. The FP8 kernels worked on GB10 even though torch doesn't list sm_121. Speed: 48 tok/s solo, 157 tok/s with 8 streams, 5.2K tok/s prompt read. Then I gave it the pagoda prompt 👇 Kolibri 1, 5 tries, thinking on: 3 blank pages from THREE.OrbitControls used the old way 1 blank page, Three.js never loaded 1 runs: floating cubes 👇 Qwen3.6 27B: first try, 12K tokens. Qwen3.8 27B: first try, 75K.

Got Aleph Alpha's new Kolibri 1 running on a single DGX Spark the day it came out. Their container is x86 only, so I built vLLM 0.29.0 (arm64) + their aleph-alpha-inference plugin. FP8 weights (74 GB) and an FP8 KV cache, 164K context. The FP8 kernels worked on GB10 even though torch doesn't list sm_121. Speed: 48 tok/s solo, 157 tok/s with 8 streams, 5.2K tok/s prompt read. Then I gave it the pagoda prompt 👇 Kolibri 1, 5 tries, thinking on: 3 blank pages from THREE.OrbitControls used the old way 1 blank page, Three.js never loaded 1 runs: floating cubes 👇 Qwen3.6 27B: first try, 12K tokens. Qwen3.8 27B: first try, 75K.

14,468 просмотров

Local models on the left. Cloud models on the right. I asked Qwen-flash-next, DeepSeek-v4.1, Fable-5.1 and Astra-6 to build an interactive volcanic island with flowing lava and boats that flee. In my opinion deepseek was the winner, and fable the disappointment Try each island yourself: Qwen · Mac Studio DeepSeek · 4× DGX Spark Fable · Cloud Astra · Cloud

Local models on the left. Cloud models on the right. I asked Qwen-flash-next, DeepSeek-v4.1, Fable-5.1 and Astra-6 to build an interactive volcanic island with flowing lava and boats that flee. In my opinion deepseek was the winner, and fable the disappointment Try each island yourself: Qwen · Mac Studio DeepSeek · 4× DGX Spark Fable · Cloud Astra · Cloud

15,234 просмотров

Qwen3.6-27B dense quality ladder. Same tasks. Same grader. Same day. How much does compression cost? FP8 - 83.3 · 29GB NVFP4 - 83.0 · ~15GB BF16 - 81.9 · 56GB GPTQ-Pro - 79.4 · 13GB IQ2_XXS - 79.0 · 9.4GB Who wins where: 🥇 Raw quality → FP8 (83.3) 🥇 Practical default → NVFP4 (83.0 at ~15GB) 🥈 Full precision → BF16 is NOT better (81.9) ❌ "Just use any 4-bit" → GPTQ-Pro loses 3.6 pts vs NVFP4 🛟 Emergency tiny → IQ2 still holds at 79.0 Overall winner: If you care about the absolute number: FP8 If you care about running it on a real box: NVFP4 is the overall pick. Same quality class as full precision. Half the size of FP8. ~4× smaller than BF16.

Qwen3.6-27B dense quality ladder. Same tasks. Same grader. Same day. How much does compression cost? FP8 - 83.3 · 29GB NVFP4 - 83.0 · ~15GB BF16 - 81.9 · 56GB GPTQ-Pro - 79.4 · 13GB IQ2_XXS - 79.0 · 9.4GB Who wins where: 🥇 Raw quality → FP8 (83.3) 🥇 Practical default → NVFP4 (83.0 at ~15GB) 🥈 Full precision → BF16 is NOT better (81.9) ❌ "Just use any 4-bit" → GPTQ-Pro loses 3.6 pts vs NVFP4 🛟 Emergency tiny → IQ2 still holds at 79.0 Overall winner: If you care about the absolute number: FP8 If you care about running it on a real box: NVFP4 is the overall pick. Same quality class as full precision. Half the size of FP8. ~4× smaller than BF16.

24,599 просмотров

This is the worst model I have ever tested, even 9b models outperform it. What even is this pagoda? Same prompt I always use. - Total time: 9 min 21 s (560.9s) - Tokens: 6,891 completion (710 reasoning + ~6,181 answer) | 84 prompt - Speed: ~12.3 tok/s end-to-end - First token: 0.15s - Output: 22,094 chars answer / 17.2KB HTML, natural stop (finish_reason=stop, no cap hit)

This is the worst model I have ever tested, even 9b models outperform it. What even is this pagoda? Same prompt I always use. - Total time: 9 min 21 s (560.9s) - Tokens: 6,891 completion (710 reasoning + ~6,181 answer) | 84 prompt - Speed: ~12.3 tok/s end-to-end - First token: 0.15s - Output: 22,094 chars answer / 17.2KB HTML, natural stop (finish_reason=stop, no cap hit)

18,306 просмотров

Videos

Anya Rossi

sweetdream.ai

SweetDream.ai•Sponsored•Livecam

Watch Anya Live

Anya is streaming live right now! Join her private show and enjoy exclusive content.

Exclusive private shows

1.2k viewers online

Private Show

Join now for exclusive access

Free preview available • Premium content

One DGX Spark. Qwen3.6-35B 64 users. 700+ tok/s. 32,768 tokens in 54 seconds. 38W Each user has their own prompt and their own KV cache, and vLLM batches every active stream through the GPU each step. Recipe → NVIDIA AI

One DGX Spark. Qwen3.6-35B 64 users. 700+ tok/s. 32,768 tokens in 54 seconds. 38W Each user has their own prompt and their own KV cache, and vLLM batches every active stream through the GPU each step. Recipe → NVIDIA AI

165,916 просмотров • 3 месяцев назад

Deepseek-v4.1 local vs Opus 5 chess: 2nd match Thinking on Reasoning low DeepSeek wins with checkmate in 22. So far thinking off Opus won Reasoning low Deepseek won Tiebreaker with thinking high underway. I’ll make a tournament style match off with more models. Best out of 3

Deepseek-v4.1 local vs Opus 5 chess: 2nd match Thinking on Reasoning low DeepSeek wins with checkmate in 22. So far thinking off Opus won Reasoning low Deepseek won Tiebreaker with thinking high underway. I’ll make a tournament style match off with more models. Best out of 3

27,700 просмотров • 1 месяц назад

Astra-6 vs DeepSeek-v4.1 cloud vs local. Same water prompt. Left → right: • Astra: “Worked for 30s” in ChatGPT • DeepSeek V4.1 Cloud: 103s, 28.5K completion tokens • DeepSeek V4.1 Local: 4× DGX Spark, 25m to HTML through Hermes 512×512 mesh. Click-to-ripple. Interacting waves. One HTML file. Might be just me but deepseek did a way better/more interesting render Astra: Cloud: Local:

Astra-6 vs DeepSeek-v4.1 cloud vs local. Same water prompt. Left → right: • Astra: “Worked for 30s” in ChatGPT • DeepSeek V4.1 Cloud: 103s, 28.5K completion tokens • DeepSeek V4.1 Local: 4× DGX Spark, 25m to HTML through Hermes 512×512 mesh. Click-to-ripple. Interacting waves. One HTML file. Might be just me but deepseek did a way better/more interesting render Astra: Cloud: Local:

22,371 просмотров • 1 месяц назад

Qwen3.8-Flash-Next (176B, NVFP4, 2× DGX Spark): 40,265 thinking tokens, 21,091 answer. Qwen3.8-27B on the same prompt: 42,988 thinking, 6,040 answer. Same question, same deliberation, 3.5× the output. 55 min vs 37 min. The 176B shipped a genuine leapfrog wave equation on a 512² lattice with reflections and interference. The 27B shipped a solid ripple tank. Both run 60fps in a single HTML file Generated on DGX Sparks on my desk.

Qwen3.8-Flash-Next (176B, NVFP4, 2× DGX Spark): 40,265 thinking tokens, 21,091 answer. Qwen3.8-27B on the same prompt: 42,988 thinking, 6,040 answer. Same question, same deliberation, 3.5× the output. 55 min vs 37 min. The 176B shipped a genuine leapfrog wave equation on a 512² lattice with reflections and interference. The 27B shipped a solid ripple tank. Both run 60fps in a single HTML file Generated on DGX Sparks on my desk.

23,168 просмотров • 2 месяцев назад

16 people chatting with Qwen3.6-35B at once ONE DGX Spark. This is a real capture, not a mockup: every token you see replays at its actual measured arrival time. Peaks ~440 tok/s total. A single user gets 105 tok/s. NVFP4 + MTP-3 on vLLM.

16 people chatting with Qwen3.6-35B at once ONE DGX Spark. This is a real capture, not a mockup: every token you see replays at its actual measured arrival time. Peaks ~440 tok/s total. A single user gets 105 tok/s. NVFP4 + MTP-3 on vLLM.

26,094 просмотров • 3 месяцев назад

Qwen3.8 Max UD-Q1_0 4× DGX Spark + llama.cpp/MTP3 53,448 tokens · 1h54m44s · 7.76 tok/s NVIDIA AI Qwen Qwen Developers

Qwen3.8 Max UD-Q1_0 4× DGX Spark + llama.cpp/MTP3 53,448 tokens · 1h54m44s · 7.76 tok/s NVIDIA AI Qwen Qwen Developers

15,149 просмотров • 2 месяцев назад

Больше нет контента для загрузки