The Open-Source AI Stack
RSS
All models

Models · qwen3

Qwen3-VL 235B-A22B Instruct

Open Alibaba · 2025-09-23 · Apache-2.0

Vision-language flagship of the Qwen3 line, 235B total weights (about 471 GB). Adds Interleaved-MRoPE for video reasoning, DeepStack multi-level ViT feature fusion, and Text-Timestamp Alignment for grounded event localization. Shipped with both Instruct and Thinking variants under Apache 2.0; native context to 256K, extensible to 1M.

Cost

$0.30 / Mtok input
$1.90 / Mtok output

· as of 2026-05-21

source ↗

Speed

50.9 tok/sec output
1170 ms TTFT

· as of 2026-05-21

source ↗

Architecture

tokens in Embedding vocab not disclosed · qwen tokenizer × N layers Attention (not disclosed) Position encoding not disclosed context 262,144 tokens MoE Router ? experts total · ? active per token Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
moe
Total params
235B
Active params
22B
Context window
262K tokens
Attention
mrope-interleaved
Position encoding
rope-interleaved
Post-training
sft, rlhf
OSI-approved
yes
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

General reasoning

MMLU-Pro 82.3 as of 2026-05-21 source ↗
GPQA-Diamond 71.2 as of 2026-05-21 source ↗

Code

LiveCodeBench 59.4 as of 2026-05-21 source ↗

Math

AIME 2025 70.7 as of 2026-05-21 source ↗

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama
AWQ Activation-aware 4-bit weight quantization for GPU serving. runs on vLLM, SGLang
MLX Apple MLX 4/8-bit layout for Apple silicon. runs on Apple MLX
FP8 8-bit float, frequently a native release on Hopper / Blackwell GPUs. runs on vLLM, SGLang, TensorRT-LLM

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · Interleaved-MRoPE for video reasoning
  • · DeepStack multi-level ViT fusion
  • · Text-Timestamp Alignment for grounded events
  • · 1M-token vision-language context

Lineage

Vision-language flagship sibling to the text Qwen3 235B MoE.

Derived from

Qwen 3 235B A22B Instruct 2025-04-28

Sources