All AI models · LLM benchmarks · Methodology

Alibaba

Qwen3.8 Flash Next

Open-weight experimental preview of the Qwen4 architecture: hybrid-attention MoE (125B total, 6B active) with vision encoder for coding, agent tasks, and image and video understanding

Model facts

Context window
262K tokens
Maximum output
262K tokens
Cheapest paid input
$0.12 per 1M tokens
Cheapest paid output
$0.4 per 1M tokens
Open weights
yes
Providers
6

Capabilities and modalities

reasoning, tool calling, structured output, attachments, temperature control, open weights, text, image, video.

Local hardware estimate

At 4K context: Q4_K_M 111.1 GiB, Q8_0 195.0 GiB, F16 352.1 GiB working memory.

Parameters
180.0 billion
Architecture
qwen4exp
Model license
other

Planning estimate adapted from the MIT-licensed whichllm estimator using metadata from Hugging Face; not a vendor minimum requirement.

Artificial Analysis benchmarks

Compare this model on the full LLM benchmark leaderboard.

ConfigurationAAICodingMathOutput tokens/s
default55.873.1not measured89.1

Observed price history

  • 2026-08-26: not published input / not published output per 1M tokens
  • 2026-08-28: $0.15 input / $0.47 output per 1M tokens
  • 2026-09-01: $0.12 input / $0.4 output per 1M tokens

API providers