Skip to content
English

Switch theme

// StepFun Step · Vision-language

GOT-OCR2_0: VRAM & local deployment

716M parameters. About 1.8 GB at 4-bit and 2.9 GB at BF16 with one 8K-token request.

released 2024-09-12licence: Commercial OKmemory figures are estimates

01Specs

PublisherStepFun · stepfun-ai
Family / seriesStepFun Step · GOT-OCR2_0
TypeVision-language (image-text-to-text)
Parameters716M Source: counted from official safetensors weights
Context length32K tokens (32,768)max_position_embeddings in config.json
Architecture24 layers · GQA
Released weightsBF16 · 1.4 GBdownload
LicenceApache-2.0 · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed.
Released2024-09-12 (Hugging Face repo created)

02How much VRAM it needs (estimate)

Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.

PrecisionWeightsTotal @ 8KTotal @ 32KSmallest GPUMac
BF16 / FP16 (original)1.4 GB 2.9 GB5.3 GB 8GB GPUMac 16GB
FP8 / INT8 / Q8_00.8 GB 2.1 GB4.6 GB 8GB GPUMac 16GB
4-bit (Q4_K_M / AWQ / GPTQ-Int4)0.4 GB 1.8 GB4.2 GB 8GB GPUMac 16GB

KV cache per request: 8K 0.8 GB · 32K 3.2 GB

Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →

03How to run it

Only runtimes the publisher or the runtime’s official library names for this model.

Which inference engine to choose →

04Official repositories & quantised builds

1 official repo(s) from stepfun-ai.

05Other StepFun Step models

All 12 →
ModelParams4-bit totalReleased
Step-5-Preview600B / 27B412 GB2026-09-20
Step-3.7-Flash198B / 11B136 GB2026-05-23
Step-3.5-Flash196B / 11B135 GB2026-02-01
Step-Audio-R1.133.5B25.0 GB2026-01-14
Step3-VL-10B10.2B8.5 GB2026-01-13
Step-Audio-EditX3.53B6.1 GB2025-10-29
Step-Audio-2-mini8.31B6.5 GB2025-08-28
step3321B / 38B247 GB2025-07-28

06Sources

Data checked 2025-02-04. Parameters, context and licence are copied from the official repository; memory figures are our estimates.

Data updated: Self-host open models(UTC+8)