Skip to content
Model Pareto

Intern-S2-Preview (35B-A3B)

Shanghai AI Lab (Intern)Open weightshuggingface.co ↗
Input
—
Output
—
Speed
~110tok/sest.
Context
262Ktokens

35B-A3B scientific multimodal model continued-pretrained from Qwen3.5. No API price found. Lab-reported scores from the HF model card results table (image; OpenCompass / VLMEvalKit, thinking mode).

Rankings

Self-hostingEstimated

How we estimate →

What it would cost to run these open weights yourself on rented GPUs. No API sells this model, so this is its price on the chart.

Hardware
1× H200 141GB
FP8 weights
Throughput
~8,390 tok/s
many requests batched
Price
$0.046–0.38 /M tok
busy → light use
On your own machine
RTX 5090 32GB (int4, 300+ tok/s single-stream)
single consumer GPU or Mac
Assumptions (5)
  • FP8 weights (35B params, 3B active per token) + 25% KV-cache headroom ≈ 44 GB
  • 1× H200 141GB at $3.59–$7.91/GPU-hour on-demand (2026-09-24)
  • ~8,390 output tok/s aggregate at batch 256 (bandwidth-bound); MoE compute scales with active params
  • Blended 3:1 input:output; prefill ~155,846 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

DeepSWE v1.1Estimated
~#43 of 46 (estimated)~35.0%
  1. #1Gemini 4 Argon78.8%
  2. #2DeepSeek V4.1 Flash 101777.0%
  3. #42Nex-N2.5-Mini36.1%
  4. ~#43Intern-S2-Preview (35B-A3B)Estimated35.0%
  5. #43Hy328.0%
No published score — estimated from related benchmarks

Writing & chat

Vision

Hard reasoning