Anro7B Real-time Chat · Median TPS ranking
Sorted by median Tokens/s, highest first. The wide bar is TPS, the thin bar TTFT; the dot colour is the chip vendor.
NVIDIA
AMD
Intel
Apple
Qualcomm
| # | Hardwarebrand · model / CPU · GPU | Median TPSTokens/s · per hardware, median across all models in the tab | AnroMarklatest assessment | Time to first tokenlower is better | Uploadsone per upload | ||
|---|---|---|---|---|---|---|---|
| 1 | Gigabyte Technology Co., Ltd. · X670 GAMING X AX V2 6 models Preliminary Anonymous · TW · AMD Ryzen 5 8600G w/ Radeon 760M Graphics · NVIDIA GeForce RTX 5090 · VRAM 32 GB · RAM 126 GB · 1,792 GB/s (spec sheet) · Plugged | 229.2TPS | … | 23 ms | 8 | ||
| 2 | AMD Ryzen 9 9950X 16-Core Processor · AMD Radeon AI PRO R9700 6 models Preliminary Archie Pixel · TW · VRAM 32 GB · RAM 190 GB · 640 GB/s (spec sheet) | 118TPS | … | 40 ms | 7 | ||
| 3 | AMD Ryzen 5 8600G w/ Radeon 760M Graphics · AMD Radeon RX 7900 XTX 6 models Preliminary Administrator · TW · VRAM 24 GB · RAM 62 GB · 960 GB/s (spec sheet) | 102.6TPS | … | 40 ms | 3 | ||
| 4 | AMD Ryzen 5 8600G w/ Radeon 760M Graphics · NVIDIA GeForce RTX 4070 Ti 6 models Preliminary Douglas on a bus · TW · VRAM 12 GB · RAM 126 GB · 504 GB/s (spec sheet) | 84.1TPS | … | 37 ms | 8 | ||
| 5 | AMD RYZEN AI MAX+ 395 w/ Radeon 8060S · AMD Radeon(TM) 8060S Graphics 6 models Preliminary Anonymous · TW · VRAM 0.5 GB · RAM 127 GB · Plugged | 37.5TPS | … | 78 ms | 3 | ||
| 6 | Micro-Star International Co., Ltd. · Creator M16 B13VE 6 models Preliminary Anonymous · TW · 13th Gen Intel(R) Core(TM) i7-13700H · NVIDIA GeForce RTX 4050 Laptop GPU · VRAM 6 GB · RAM 32 GB · 216 GB/s (spec sheet) · Plugged | 23.4TPS | … | 285 ms | 6 | ||
| 7 | ASUSTeK COMPUTER INC. · ROG Strix G614JU_G614JU 6 models Preliminary Hoshikawa Rio · JP · 13th Gen Intel(R) Core(TM) i9-13980HX · NVIDIA GeForce RTX 4050 Laptop GPU · VRAM 6 GB · RAM 96 GB · 216 GB/s (spec sheet) · Plugged | 22.4TPS | … | 256 ms | 6 |
測試標準與標記規則Test standard & badge policy
硬體自選比對Hardware Side-by-Side Comparison
挑選 2 到 4 款硬體，水平對齊規格與各分頁的中位數 TPS、TTFT 與並發席位。Pick 2 to 4 hardware items and align specs plus median TPS, TTFT and seats per tab.
選型矩陣 · 硬體 × 模型大小Sizing matrix · hardware × model size
一列一台硬體、一欄一個模型尺寸，格子是該尺寸分頁的中位數 TPS。橫著看：這台機器跑得動多大的模型；先選尺寸再直著看：哪台機器跑得動它。One row per machine, one column per model size; each cell is that tab's median TPS. Read across for how big a model a machine can run; pick a size and read down for which machines can run it.
AnroMark 地端 AI 算力天梯圖中 · 每一列是一種硬體規格（CPU、GPU、記憶體相同的機器合併計算），TPS 為該規格在此分頁所有模型測試的中位數；一筆上傳算一筆，未滿 10 筆標示「初步樣本」，達 10 筆標示「社群驗證」。測試次數包含同一筆上傳內的重複跑測。TTFT 越低越好。AnroMark 是該硬體最新一筆評估的本機總分，跨模型尺寸加權，各分頁相同。AnroMark Local AI Ladder · Each row is one hardware configuration (machines with the same CPU, GPU and memory are merged); TPS is the median across all model runs of that hardware in this tab. The uploads column counts one per upload (< 10 → Preliminary, ≥ 10 → Community), while run counts include repeated runs within one assessment. Lower TTFT is better. AnroMark is the machine score of the hardware's latest assessment, weighted across model sizes and the same in every tab.
AnroMark 計分算法
每個模型
- 分數 = TPS 70% + TTFT 30%，再乘 CPU 折扣和長提示折扣。
- 滿標 TPS = 目標 × 1.25。目標凍結在 2026.1：3B 是 100 TPS / 0.1 秒，14B 是 50 / 0.45 秒，20B 是 40 / 0.8 秒。
- 未達標：TPS 用曲線 ratio^1.9。超過達標：TPS 線性往上加，沒有上限。
- TTFT 最快只算到目標，更快不再加分。
- 純 CPU 約打 87 折。gpt-oss:20b 長提示沒跑完，再乘 0.5–1.0。
總分
- Small（小於 4.5B）20%，Mid（4.5–16B）40%，High（大於 16B）40%。
- 同一層有多個模型就均分。沒測的層拿掉，其餘權重重新加總成 100%。
- 已選但失敗：權重保留，該模型 0 分。
- 總分 = 各模型分數 × 權重。沒有 1000 上限，也沒有 950–999 菁英映射。
- 只測小模型時，小層變成 100%，高 TPS 可以超過 2000，不代表完整 20B 機器。
- 各模型都是 2 倍達標 TPS、TTFT 也滿標時，總分約 1700。
評等
How AnroMark is scored
Per model
- Score = TPS 70% + TTFT 30%, then multiplied by the CPU-only and long-prompt discounts.
- Full-marks TPS = target × 1.25. Targets are frozen at 2026.1: 3B is 100 TPS / 0.1 s, 14B is 50 / 0.45 s, 20B is 40 / 0.8 s.
- Below target: TPS follows the curve ratio^1.9. Above target: TPS adds linearly, with no cap.
- TTFT counts down to the target at most; faster earns nothing more.
- CPU-only runs take roughly a 13% cut. If gpt-oss:20b did not finish the long prompt, a further 0.5–1.0 multiplier applies.
Total
- Small (under 4.5B) 20%, Mid (4.5–16B) 40%, High (over 16B) 40%.
- Several models in one tier share its weight equally. Untested tiers are dropped and the remaining weights re-summed to 100%.
- A selected model that failed keeps its weight and scores 0.
- Total = Σ model score × weight. No 1000 cap and no 950–999 elite mapping.
- With only small models tested, the small tier becomes 100%; a high TPS can exceed 2000 and does not stand for a full 20B machine.
- With every model at 2× full-marks TPS and TTFT on target, the total is about 1700.
Rating
anrotec.ai · anromark.ai · 加入 AnroMarkJoin AnroMark · 授權合約EULA · 隱私權政策Privacy Policy · © Anrotec Inc.