Model
Model explorer

Skywork-R1V (38B)

OPEN
Kunlun Tech · Skywork R1V family · released Mar 18, 2025

First open-sourced multimodal reasoning model with visual chain-of-thought, pairing InternViT-6B with DeepSeek-R1-Distill-Qwen-32B

ReasoningCodingVisionFunction callingTool useAgentic
1280.5
Elo · rank #174
Parameters
38B
Active params
38B (dense)
Context
128K tokens
Architecture
InternViT-6B-448px-V2_5 vision encoder + DeepSeek-R1-Distill-Qwen-32B LLM via lightweight visual projector
License
MIT
Languages
—
API price (in/out)
No hosted API
Modalities
text · vision
Benchmark results
Bar shows position within the tracked field; marker = field best
AIMEMath72.0%#101
best: GPT-5.2 · 100.0%
GPQA DiamondReasoning61.6%#167
best: GPT-6 Astra · 96.0%
MATH-500Math94.0%#39
best: GPT-5 · 99.4%
MathVistaVision67.5%#34
best: Seed 2.1 Pro · 90.7%
MMMUVision69.0%#59
best: Claude Fable 5 · 89.3%
Run it locally
VRAM @ Q4
—
VRAM @ FP16
76 GB
Fits on (Q4)
Multi-node cluster required
Throughput data unavailable.
Quantizations
—
Fine-tune it
Permissive
QLoRA25.9 GB1× RTX 5090 32GB
LoRA81.0 GB1× H200 141GB
Full fine-tune610.0 GB4× B200 192GB
QLoRA SFT on ~10k samples ≈ $18.53 (1× RTX 5090 32GB)
Plan a fine-tune →
Skywork R1V family
Elo progression across releases
Skywork-R1V (38B)Mar 20251280.5
Resources
HFSkywork/Skywork-R1V-38BGHSkyworkAI/Skywork-R1V
API price weights · each benchmark row carries its own source badge (see methodology)