RTX Spark評測:74B AI模型裝進筆電,官方沒附跑分 | RTX Spark Review: 74B AI Model, No Benchmarks Yet
By Kit 小克 | AI Tool Observer | 2026-10-16
🇹🇼 RTX Spark評測:74B AI模型裝進筆電,官方沒附跑分
RTX Spark今天(10/16)正式開賣,這是Nvidia第一款把資料中心級Grace CPU + Blackwell GPU塞進Windows筆電的產品,同步搭配新開源模型Nemotron 3.5 Bolt 74B,主打「74B參數模型塞進24GB記憶體,筆電就能跑AI Agent」。聽起來很猛,但細看規格表會發現官方跳過了一件最重要的事:跑分。
RTX Spark硬體規格:128GB統一記憶體,要價2599起
RTX Spark搭載20核Grace CPU + 6144核Blackwell GPU,記憶體最高到128GB,宣稱FP4精度下可達1 petaflop算力,功耗45W到80W之間。首發機型包括Microsoft Surface Laptop Ultra、HP OmniBook Ultra 16、ASUS ProArt系列。
- 入門版:64GB記憶體,約2599美元
- 頂規128GB版本:依品牌不同落在3500~6799美元
- 記憶體頻寬300GB/s,比桌機獨顯慢不少
Nemotron 3.5 Bolt 74B:MoE架構,實際只用9B參數運算
Nemotron 3.5 Bolt是混合專家(MoE)模型,總參數74B,但每次推論只啟用9B,再用2-bit量化塞進24GB,讓一般筆電也能本機跑。相容llama.cpp、Ollama、vLLM、LM Studio,Nvidia官方宣稱它是同量級開源模型中PinchBench排名第一、BrowseComp排名第二。
誠實說:沒有跑分,沒有tokens/sec,就是最大問題
這才是重點。Nvidia這次發布稿寫滿記憶體體積和理論算力,卻完全沒附上速度數字(tokens/sec)、沒有第三方獨立評測、所謂的排名也只是單方面宣稱。換句話說,Nemotron 3.5 Bolt目前看起來像是「規格很漂亮,實際好不好用沒人驗證過」的狀態。對照RTX Spark的記憶體頻寬只有300GB/s(遠低於桌機RTX顯卡),實際推論速度很可能比官方行銷話術保守許多。
值不值得買:創作者、本機AI愛好者可以關注,一般人先觀望
如果你是做本機AI Agent開發、不想把資料送上雲端、又想跑中大型開源模型,RTX Spark + Nemotron 3.5 Bolt的組合值得實測看看。但2599美元起(頂規逼近7000美元)不是衝動購物的價位,建議等第三方跑分和實測影片出來再下手,別只看官方新聞稿就掏錢。
好不好用,試了才知道。
🇺🇸 RTX Spark Review: 74B AI Model, No Benchmarks Yet
RTX Spark laptops officially went on sale today (Oct 16), Nvidia's first attempt at squeezing a data-center-class Grace CPU + Blackwell GPU into a Windows laptop. It launches alongside a new open model, Nemotron 3.5 Bolt 74B, pitched as "a 74B-parameter agent model that fits in 24GB of memory, runnable on a laptop." Sounds impressive — until you notice what's missing from the spec sheet: actual benchmarks.
RTX Spark Hardware: 128GB Unified Memory, Starting at $2,599
RTX Spark pairs a 20-core Grace CPU with a 6,144-core Blackwell GPU, up to 128GB of unified memory, and a claimed 1 petaflop of FP4 compute, running at 45W-80W. Launch devices include the Microsoft Surface Laptop Ultra, HP OmniBook Ultra 16, and ASUS ProArt series.
- Entry config: 64GB memory, around $2,599
- Top-tier 128GB config: $3,500–$6,799 depending on brand
- Memory bandwidth is 300GB/s — notably slower than a desktop discrete GPU
Nemotron 3.5 Bolt 74B: MoE Architecture, Only 9B Active Parameters
Nemotron 3.5 Bolt is a mixture-of-experts model with 74B total parameters but only 9B active per token, 2-bit quantized to fit in 24GB so it runs locally on consumer hardware. It works with llama.cpp, Ollama, vLLM, and LM Studio, and Nvidia claims it ranks #1 among open models its size on PinchBench and #2 on BrowseComp.
The Honest Problem: No Benchmarks, No Tokens/Sec, No Independent Testing
This is the part that matters. Nvidia's announcement is packed with memory footprint and theoretical FLOPS numbers, but there's no published speed (tokens/sec), no independent third-party evaluation, and the ranking claims come from Nvidia alone. In short, Nemotron 3.5 Bolt currently looks like "great specs, unverified real-world performance." Combined with RTX Spark's relatively modest 300GB/s memory bandwidth, actual inference speed could fall well short of the marketing copy.
Is It Worth Buying? Wait for Real Reviews First
If you're building local AI agents, want to keep data off the cloud, and want to run mid-to-large open models, the RTX Spark + Nemotron 3.5 Bolt combo is worth watching. But at $2,599 and up (nearly $7,000 for the top config), this isn't an impulse buy — wait for independent benchmarks and hands-on reviews before trusting the press release.
You won't know until you try it.
Sources / 資料來源
- NVIDIA Expands Nemotron Family With Super VL and Bolt Models
- NVIDIA RTX Spark laptops launch October 16, prices start at $2,599 - VideoCardz
- Nvidia's Open Nemotron 3.5 Puts a 74B Agent Model in 24GB - FourWeekMBA
延伸閱讀 / Related Articles
- Claude假線報評測:AI誤報兇殺案,Anthropic緊急斷網 | Claude Fake Tip Review: AI Sent False Murder Tip to Police
- XBOW評測:GPT-5.6挖出Chrome零日,奪25萬美元賞金 | XBOW Review: GPT-5.6 Finds Chrome Zero-Day, Wins $250K
- LiteLLM漏洞評測:AI閘道曝admin後門,近億下載受影響 | LiteLLM Vulnerability Review: AI Gateway's Admin Backdoor
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言