Wink Pings

Baidu ERNIE 5.5 Flash scores 1668, and the 280B model is called Flash

On the Code Arena: WebDev leaderboard, ERNIE 5.5 Flash Preview ranks second among open-source models with 1668 points. 280B total parameters, 14B active, 1M context, Apache 2.0 weights to be released on November 25. The comments are discussing more than just the score.

Arena.ai's Code Arena: WebDev leaderboard has been updated. Baidu's ERNIE 5.5 Flash Preview scored 1668, ranking second among open-source models.

Where the score sits:

  • Above DeepSeek-V4.1-Flash's 1664
  • Above Kimi K3's 1658
  • Roughly tied with Qwen3.8-Max's 1670
  • Counting closed-source models as well, it ranks 17th

Arena.ai specifically notes that this score comes from real task performance, not static benchmark scores.

image

The model specs are also noteworthy:

  • 280B total parameters, 14B activated per token
  • 1M token context window
  • Generation speed of 200+ tokens/s
  • Weights open-sourced under Apache 2.0, releasing on November 25

image

A 280B model called Flash. Someone in the comments could not resist: with a name like that, the flash feels a bit heavy. Flash usually brings to mind small models; putting 280B under this name creates a kind of contrast.

Another detail comes from coldtrace's observation. The HuggingFace 30-day download counts for the top ten on the WebDev leaderboard are inverted relative to Elo scores:

  • Kimi-K3.1-Pro: 33.6K downloads
  • DeepSeek-V4.1-Flash: 348.9K
  • Qwen3.8-Flash-Next-GGUF: 1.97M

A 40-point Elo gap corresponds to tens of times difference in downloads. Day-1 vLLM/GGUF quantization support may have a bigger impact on actual use than leaderboard ranking.

image

Back to ERNIE 5.5 Flash. The 14B active parameter count is the key number. For a 280B MoE model, only 14B are activated per inference, meaning mid-sized teams have a chance to deploy it themselves. If the Apache 2.0 weights are released on time, API costs will get a fresh round of recalculation. Someone mentioned in the comments: there is not a single US model in the top 15.

image

The weights have not been released yet. Before November 25, 1668 is a number on the leaderboard, and Apache 2.0 is a promise. If you want to try it early, there is a free version available to run on OpenRouter: https://openrouter.ai/baidu/ernie-5.5-flash-preview:free

Leaderboard link: https://arena.ai/leaderboard/code/webdev

发布时间: 2026-11-12 17:28