Skip to content
Model Pareto

Nex-N2.5-Pro

Nex AGIOpen weightshuggingface.co ↗
Input
$0.075/M tok
Output
$0.25/M tok
Blended
$0.119/M tok · 3:1
Speed
~90tok/sest.
Context
262Ktokens
Cost per task
~$0.10/taskest.

397B-A17B MoE (Qwen3.5-MoE arch, config.json: 512 experts, 10 active/token + shared expert). Weights were marked "coming soon" in the README at the Sep 8 2026 launch but are live on HF now (checked 2026-09-28, last modified 2026-09-17). Ships natively quantized in FP8 (config.json quantization_config: format "float-quantized", FP8_BLOCK).

Rankings

Self-hostingEstimated

How we estimate →

What it would cost to run these open weights yourself on rented GPUs, compared with the API price above.

Hardware
4× B200 180GB
FP8 weights
Throughput
~12,969 tok/s
many requests batched
Price
$0.20–1.76 /M tok
busy → light use
On your own machine
Mac Studio M5 Ultra 512GB (int4, ~99 tok/s single-stream)
single consumer GPU or Mac
Assumptions (5)
  • FP8 weights (397B params, 17B active per token) + 25% KV-cache headroom ≈ 496 GB
  • 4× B200 180GB at $5.98–$14.24/GPU-hour on-demand (2026-09-24)
  • ~12,969 output tok/s aggregate at batch 371 (bandwidth-bound); MoE compute scales with active params
  • Blended 3:1 input:output; prefill ~250,147 tok/s
  • Low = 75% utilization at the low GPU price; high = 20% at the high price

Benchmark scores

~ italic, dashed = no published score yet, estimated from related benchmarks.

Coding

~#3 of 21 (estimated)~83.6%
  1. #1Claude Opus 596.0%
  2. #2Claude Sonnet 585.2%
  3. ~#3Nex-N2.5-ProEstimated83.6%
  4. #3Beam 1.0 501B A23B81.6%
No published score — estimated from related benchmarks
~#26 of 102 (estimated)~26.6%
  1. #1Kimi K3.1 Pro67.2%
  2. #2Claude Sonnet 5.563.6%
  3. #24Mistral Large 426.8%
  4. ~#26Nex-N2.5-ProEstimated26.6%
  5. #26Grok 4.725.8%
No published score — estimated from related benchmarks
LiveCodeBenchEstimated
~#6 of 28 (estimated)~88.9%
  1. #1Qwen3.8-Omni-Flash92.6%
  2. #2Qwen3.8 Flash-Next91.9%
  3. #5Nemotron 3 Ultra89.0%
  4. ~#6Nex-N2.5-ProEstimated88.9%
  5. #6Beam 1.0 501B A23B88.3%
No published score — estimated from related benchmarks
SciCodeEstimated
~#23 of 119 (estimated)~55.9%
  1. #1Claude Fable 5.267.7%
  2. #2Claude Opus 5.566.9%
  3. #22Claude Opus 556.4%
  4. ~#23Nex-N2.5-ProEstimated55.9%
  5. #23DeepSeek V4.1 Flash 101755.9%
No published score — estimated from related benchmarks
~#14 of 26 (estimated)~47.1
  1. #1Claude Opus 5.561.7
  2. #2Claude Opus 559.7
  3. #13DeepSeek V4.1 Flash47.3
  4. ~#14Nex-N2.5-ProEstimated47.1
  5. #14Grok 4.647.0
No published score — estimated from related benchmarks

Agents

Not comparable across labs (1)

Kept for reference, not counted in rankings: these use a lab-specific task set, answer key or scoring, so scores can't be compared fairly across labs. Only published scores are shown.