Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Čtvrtkon #143 - Local AI (na Macu) micro talk

Sponsored · Ship Features Fearlessly Turn features on and off without deploys. Used by thousands of Ruby developers. →

Čtvrtkon #143 - Local AI (na Macu) micro talk

Martin Hlaváč

October 15, 2026

Transcript

18 slides
  1. Hardware Zařízení Paměť Propustnost ↑ Cena NVIDIA DGX Spark 128

    GB 273 GB/s ~170 000 Kč MacBook Pro 14” Apple M5 Pro 64 GB 307 GB/s ~105 000 Kč MacBook Pro 16” Apple M5 Max 128 GB 614 GB/s ~203 000 Kč Mac Studio Apple M5 Ultra 256 GB 1,2 TB/s ~329 000 Kč NVIDIA RTX PRO 6000 96 GB 1,79 TB/s ~420 000 Kč
  2. Model Parametry BF16 ↑ Q8 Q4 Qwen3.8-27B 27B ~54 GB

    ~29 GB ~18 GB Muse Glimmer 30B 30B ~60 GB ~32 GB ~17 GB Qwen3.6-35B-A3B 35B / 3B active ~70 GB ~37 GB ~21 GB Qwen3.8-Flash-Next 180B / 6B active ~354 GB ~188 GB ~120 GB GLM-5.3-Flash 320B / 18B active ~640 GB ~320 GB ~201 GB MiniMax M3 428B / 23B active ~856 GB ~454 GB ~261 GB DeepSeek V4.1-Flash 763B / 8-16B active ~1,5 TB ~800 GB ~445 GB Kimi K3 2,8T / 104B active ~5,6 TB ~3,0 TB ~1,56 TB
  3. Model na 64 GB Parametry BF16 Q8 Q4 Qwen3.8-27B 27B

    🟡 🟢 🟢 Muse Glimmer 30B 30B 🔴 🟢 🟢 Qwen3.6-35B-A3B 35B / 3B active 🔴 🟢 🟢 Qwen3.8-Flash-Next 180B / 6B active 🔴 🔴 🔴 GLM-5.3-Flash 320B / 18B active 🔴 🔴 🔴 MiniMax M3 428B / 23B active 🔴 🔴 🔴 DeepSeek V4.1-Flash 763B / 8-16B active 🔴 🔴 🔴 Kimi K3 2,8T / 104B active 🔴 🔴 🔴
  4. Model na 128 GB Parametry BF16 Q8 Q4 Qwen3.8-27B 27B

    🟢 🟢 🟢 Muse Glimmer 30B 30B 🟢 🟢 🟢 Qwen3.6-35B-A3B 35B / 3B active 🟢 🟢 🟢 Qwen3.8-Flash-Next 180B / 6B active 🔴 🔴 🔴 GLM-5.3-Flash 320B / 18B active 🔴 🔴 🔴 MiniMax M3 428B / 23B active 🔴 🔴 🔴 DeepSeek V4.1-Flash 763B / 8-16B active 🔴 🔴 🔴 Kimi K3 2,8T / 104B active 🔴 🔴 🔴
  5. Model na 256 GB Parametry BF16 Q8 Q4 Qwen3.8-27B 27B

    🟢 🟢 🟢 Muse Glimmer 30B 30B 🟢 🟢 🟢 Qwen3.6-35B-A3B 35B / 3B active 🟢 🟢 🟢 Qwen3.8-Flash-Next 180B / 6B active 🔴 🟢 🟢 GLM-5.3-Flash 320B / 18B active 🔴 🔴 🟡 MiniMax M3 428B / 23B active 🔴 🔴 🔴 DeepSeek V4.1-Flash 763B / 8-16B active 🔴 🔴 🔴 Kimi K3 2,8T / 104B active 🔴 🔴 🔴
  6. DeepSeek V4 Flash (0731) V4.1 Flash Stáří (k 15. 10.

    2026) 76 dní 35 dní AA Intelligence Index 34 39 (+5) Terminal-Bench 4.0 12 % 27 % (2,3×) Parametry celkem 304B 763B (2,5×) BF16 ~608 GB ~1,5 TB (2,5×) Q8 ~300 GB ~800 GB (2,7×) Q4 ~155 GB ~445 GB (2,9×)
  7. Lokální AI stack → Model ◦ Qwen → Runtime /

    Server ◦ Splash (MLX) / LM Studio → API ◦ OpenAI-compatible → Agent ◦ OpenCode → Tools/MCP
  8. Lokální ≠ offline Web search Cena Limit Exa AI Zdarma

    s limitem ~1400 Tavily Zdarma s limitem 1000/500 Brave Search Kredit zdarma ~1000 SoloCrawl Zdarma lokálně ∞*
  9. Vyplatí se to? Lokálně Cloud Soukromí Frontier kvalita Fixní náklady

    Pay per use Může běžet offline Dynamický výkon Plná kontrola Bez údržby

More Decks by Martin Hlaváč

Other Decks in Programming