AIAI Tech Engine
Home/Benchmarks/DGX Spark vs M1 Max vs M2 Max: same REV-441 needle, same Angry Birds HTML | AI Tech Engine
#Qwen3.8#Ornith#DGX Spark#M2 Max#M1 Max#needle#Angry Birds#실측

DGX Spark vs M1 Max vs M2 Max: same REV-441 needle, same Angry Birds HTML | AI Tech Engine

Same REV-441 needle (~2.5k–82k) and Angry Birds-style HTML on DGX Spark Qwen3.8-27B-NVFP4, M1 Max (prior post), and M2 Max Ornith/Qwen. Spark 82k TTFT 4.1s. All needles matched. GIFs included.

AIAI Tech Engine Lab
Independent lab measurement
SPONSORED / ADVERTISEMENT
예약된 광고 영역

English summary page — full tables in Korean version. Same REV-441 needle (~2.5k–82k) and Angry Birds-style HTML on DGX Spark Qwen3.8-27B-NVFP4, M1 Max (prior post), and M2 Max Ornith/Qwen. Spark 82k TTFT 4.1s. All needles matched. GIFs included. 한국어 전문 →

핵심 결론 (Bottom Line) 2026-08-31, 같은 REV-441 needle과 같은 Angry Birds형 단일 HTML 프로토콜을 세 박스에 넣었다. DGX Spark + Qwen3.8-27B-NVFP4 + DFlash2의 ~82k TTFT는 4.1s(M1 MTPLX Qwen 18분, M1 omlx Qwen ~31분). M2 Max Ornith oQ4e-mtp의 ~41k TTFT는 6.9s(M1 Ornith 117s). M2 Qwen3.8-27B-Uncensored-oQ4e-mtp는 short decode ~30 tok/s, Angry Birds 7000은 잘리고 12000에서 완결. needle은 잰 구간 전부 맞춤. 용어는 어색하면 영어 유지: needle, prefill, decode, TTFT, tok/s.

이전 글(M1 Max만): M1 Max 64GB 실측. 오늘은 그 칸을 Spark·M2로 채운다.

장비

M1 Max 64GB (2026-08-30)M2 Max 96GB (오늘)DGX Spark GB10 (오늘)
런타임MTPLX / omlxomlx :8000SGLang + DFlash2 K=8 :8001
모델Qwen3.8-27B FP16 MTP · 4bit+DFlash · Ornith 4bitOrnith-1.5-35B-A3B-oQ4e-mtp · Qwen3.8-27B-Uncensored-oQ4e-mtpqwen38-27b-stock-nvfp4
thinkingoffoffoff

참고: M2의 Qwen3.8-27B-DFlash2rope_theta 로드 실패로 오늘 표에서 빠졌다. Qwen 칸은 Uncensored-oQ4e-mtp(ctx 32k → 41k/82k 미실시).

같은 needle, 같은 토큰

가짜 billing 모듈 중간 settle_invoice(... tax_rate=0.08) + docstring REV-441 / 0.10. JSON만. 런마다 RUN_ID/CACHE_BUST.

contextM1 MTPLX QwenM1 omlx Qwen DFlashM1 OrnithSpark Qwen NVFP4M2 OrnithM2 Qwen oQ4e
~2.5kTTFT 22.1s · decode 20.2 · 맞춤wall ~51.5s · 맞춤TTFT 6.6s · decode 92.7 · 맞춤TTFT 0.55s · decode 66.3 · 맞춤TTFT 3.32s · decode 91.9 · 맞춤TTFT 19.6s · decode 30.6 · 맞춤
~13.5kTTFT 73.6s · decode 20.3 · 맞춤wall ~275s · 맞춤TTFT 33.2s · decode 73.6 · 맞춤TTFT 0.23s · decode 60.2 · 맞춤TTFT 3.90s · decode 82.1 · 맞춤TTFT 19.4s · decode 28.1 · 맞춤
~41kTTFT 414s · decode 13.9 · 맞춤wall 783s · 맞춤TTFT 117s · decode 58.8 · 맞춤TTFT 0.33s · decode 63.7 · 맞춤TTFT 6.85s · decode 66.0 · 맞춤ctx 32k — 미실시
~82kTTFT 1091s · decode 8.5 · 맞춤wall 1863s · 맞춤TTFT 305s · decode 41.6 · 맞춤TTFT 4.06s · decode 58.2 · 맞춤ctx 65k — 미실시미실시

Spark prompt_tokens: 2531 / 14126 / 42812 / 85440. 맞추느냐는 갈리지 않고, TTFT/prefill이 갈린다.

짧은 생성

과제M1 MTPLX QwenM1 OrnithSpark QwenM2 OrnithM2 Qwen oQ4e
merge_intervals29.764.573.384.529.8
two_sum32.077.175.880.130.8
해시테이블 산문19.464.829.170.222.1

같은 프롬프트, Angry Birds형 HTML

960×540 canvas, slingshot, bird 3 · wood 4 · pig 3, R restart. thinking off, 온도 0.4. 첫 샷 max_tokens=7000, 이어서 12000.

M1 OrnithM1 omlx QwenM2 OrnithM2 Qwen oQ4eSpark Qwen
7000 wall / decode120s / 58.91698s / 4.292.8s / 76.3265s / 26.6155s / 45.3
7000 결과발사(루프 닫음)HUD만거의 완결→close잘림잘림
1200096s / 76.7 · 완결322s / 27.2 · 완결221s / 45.4 · 완결

GIF

M1 Ornith (원문 컷):

M1 Max Ornith slingshot

M2 Ornith 7000 close:

M2 Max Ornith slingshot

DGX Spark Qwen 12000 완결:

DGX Spark Qwen slingshot

M2 Qwen 12000 완결 GIF: /images/m2-max-qwen38-slingshot-12k-2026-08-31.gif

Same prompt, 3D architecture HTML (Three.js)

Same single-HTML protocol as Angry Birds. Subject: modern museum / pavilion, Three.js r160 CDN, auto orbit, dusk lighting. max_tokens=12000, thinking off. GIFs = 6s Playwright orbit.

Note: M1 MacBook was offline; M1-post Ornith 4bit weights were re-run on M2 Studio omlx (artifact compare). Decode numbers are M2 Studio.

M2 Max · Ornith oQ4e-mtp · open HTML
M2 Ornith 3D pavilion

~88s · ~76 tok/s · 6526 tok — Barcelona pavilion vibe, warm sky, works.

M1 Max · Ornith 4bit (M1 post model) · open HTML
M1 Ornith 3D pavilion

~75s · ~74 tok/s · 5453 tok — curtain wall + sunset, playable.

VerdictM2 Ornith oQ4e-mtpM1 Ornith 4bit
Browser renderyesyes
Architecture polish1st2nd
vs DGX Spark trioFlash-Next > Qwen27B ≫ DeepSeek — both Mac Ornith beat Spark DeepSeek badly.

무엇을 고를 것인가

  • 같은 27B로 긴 컨텍스트를 빨리 → DGX Spark Qwen NVFP4 + DFlash2
  • Mac에서 긴 HTML / 빠른 MoE → M2 Ornith oQ4e-mtp (M1 Ornith보다 TTFT·decode 유리)
  • Mac Qwen3.8-27B만 → M1 기준 MTPLX ≫ omlx DFlash. M2 Qwen oQ4e는 short ~30 tok/s, 장문맥은 Spark에 못 미침

하지 않은 일

  • M2 Qwen DFlash2 로드 수정 후 41k/82k
  • Spark Flash-Next / DeepSeek-v4-Flash → 새 글 (모델 탭)
  • AMD AI Max
  • 131k needle on Spark