Baidu's ERNIE 5.5 Flash Preview is the lab's first post-trained checkpoint on ERNIE 5.5 Flash Base — still a 280B MoE with 14B active parameters and 1M-token context, but tuned for coding, 3D, knowledge work, and video pipelines inside agent harnesses.
This week, Arena-linked social posts say it is trending toward #2 among open models on Code Arena: WebDev at 1668 points. That headline is spreading faster than the license tag on the actual leaderboard row.
This post separates what Baidu's release page confirms, what Arena publishes, and what is still a promise.
TL;DR: claimed vs verified
| Claim | Verified? | Source |
|---|---|---|
| Preview API live on Qianfan | Yes | Baidu ERNIE release page |
| 280B total / 14B active / 1M context | Yes | Same release page |
| Apache 2.0 weights soon | Promised, not shipped | Release page wording |
| 1668 WebDev Arena score | Reported on Arena row | Arena WebDev snapshots cited in third-party analysis |
| #2 open model on WebDev | Not verified as open | Row labeled Proprietary with AutoEval, N/A votes in public tables |
| Beats DeepSeek-V4.1-Flash (1664) and Kimi K3 (1658) | Arithmetic only | Error bars overlap on the Preview row |
What Baidu actually shipped
The November 11, 2026 release page is a product checkpoint, not another base dump.
Baidu states:
- ERNIE 5.5 Flash Preview is the first post-trained model on ERNIE 5.5 Flash Base.
- Strengths listed: coding, 3D modeling, knowledge work, content and video creation.
- API is available on Qianfan.
- Weights will be open-sourced under Apache 2.0 soon.
The page is heavy on demo videos — Unreal Wuxia game builds, Blender airports, CAD engines, symphony audio, and agent-produced research PDFs. Those are marketing-grade capability proofs, not independent benchmarks.
Critically, the release page does not embed the 1668 WebDev figure in the HTML we fetched. That number arrives through Arena's ecosystem and reposts.
Training is ongoing. Baidu warns Preview may fall short on some skills until the official ERNIE 5.5 Flash finish line.
The WebDev Arena row everyone is quoting
Code Arena WebDev measures agentic front-end development — multi-step coding with tools, not chat helpfulness.
Third-party analysis of the November 10–12, 2026 board (summarized in community posts) describes the ERNIE 5.5 Flash Preview row roughly as:
- Score: 1668 published as 1668 +16/-16
- Rank: N/A displayed between numeric ranks
- Votes: N/A human vote count
- Label: AutoEval (Arena's own evaluation instrument)
- License tag on row: Proprietary
- Listed price: on the order of $0.28 in / $1.13 out per million tokens on Qianfan pricing mirrors
Meanwhile, DeepSeek-V4.1-Flash and Kimi K3 rows on the same board carry MIT or vendor open licenses with tens of thousands of human votes in snapshots that circulated this week.
So the sentence "#2 open model on WebDev" bundles three different things:
- A real score on a real benchmark family
- A license status that is not open yet on the row
- A ranking comparison that mixes AutoEval with human-voted MIT models
That is why explainx.ai treats this as claimed vs verified, not a launch recap.
Why the error bar matters
If the Preview interval is roughly 1652–1684, then DeepSeek at 1664 and Kimi at 1658 sit inside the band.
Headlines that say Preview is "just above" those models state point estimates only. They are not significance-tested victories on the AutoEval row.
For builders choosing a web agent stack, treat 1668 as "worth a trial API key", not "dominates open WebDev."
Open weights: Base is out, Preview is not
We already documented ERNIE-5.5-Flash-Base on Hugging Face with vLLM/SGLang serve examples.
Preview is the post-trained product slice. Until Apache 2.0 weights land:
- Reproducibility stays on Baidu's API
- Arena rows may stay Proprietary
- Comparisons to MIT DeepSeek or Qwen self-host paths are unequal
When weights ship, expect a new post or update on serve recipes — do not assume Preview borrows Base's FP8/BF16 cards without a new card.
How WebDev differs from generic coding benches
WebDev Arena targets design-to-site and reference-based front-end workflows — closer to product engineer tasks than HumanEval snippets.
That is why a MoE preview can rank well while your local CLI codegen model feels unrelated. Match the bench to your job:
| If you build… | WebDev Arena signal | Better cross-check |
|---|---|---|
| Marketing landing pages from Figma | High | Run your own reference designs |
| Backend APIs only | Low | Use coding harness evals |
| Long-horizon game tooling | Medium | Baidu's UE demo videos, not Arena |
See also Baidu OCR document models for a different Baidu open-weight lane.
API access and cost posture
Baidu routes Preview through Qianfan. Arena-linked pricing mirrors near 28 cents per million input tokens and about $1.13 per million output in third-party tables — confirm on your Qianfan console before you budget agent loops.
Contrast with open MIT stacks you host yourself: you pay GPUs, not per-token API, but you carry ops.
Web development agent workflow across multi-step frontend coding tasks
What social posts get right and wrong
Right:
- Baidu did ship a named Preview with heavy agent demos.
- Arena does show a 1668 WebDev score row for the Preview name.
- 280B / 14B active matches the official release text.
Wrong or premature:
- Calling it "#2 open model" before Apache 2.0 weights and before open-license row tags align.
- Treating AutoEval N/A votes like 13k-vote MIT DeepSeek rows.
- Ignoring overlapping error bars when comparing rivals.
Arena's "new release" rail versus the main board
Arena's homepage New Release Rankings strip also highlights Ernie 5.5 Flash Preview around #12 in WebDev with a Preview tag. That rail is a discovery widget, not a substitute for reading license, vote count, and evaluation mode on the full table.
When marketing teams screenshot the rail, they often crop out Proprietary badges. Procurement teams should not.
AutoEval rows exist so labs can publish controlled scores before large human battles complete. They are useful early signal. They are not interchangeable with MIT-licensed models that accumulated thousands of human preferences.
Qianfan API versus self-host Base weights
Until Preview weights arrive, you face a fork:
| Path | You get | You accept |
|---|---|---|
| Qianfan Preview API | Post-trained behaviors and demos | Per-token spend, data handling policy |
| Self-host Flash Base | Open weights today | Fine-tune or instruct-tune yourself |
| Wait for Apache Preview | Parity with marketing | Unknown ship date |
Teams building WebDev agents should not pretend the Base checkpoint is a drop-in replacement for Preview. Base completes text; Preview is already positioned for long-horizon agent harnesses on Baidu's page.
What to do this week
- Read Baidu's release page for API scope — not for WebDev rank proof.
- Open Arena WebDev and inspect license, votes, and AutoEval columns yourself.
- If you need open weights today, start from Flash Base and fine-tune — different artifact.
- Run one reference-design task on Qianfan Preview and one on your incumbent MIT model; log cost, time, and diff quality.
- Watch for HF repo + license file — that is the event that flips this from claimed vs verified to a standard open launch post.
How explainx.ai will update this thread
We will update the Flash Base post when Preview weights mirror the Apache promise, and we will revise this article if Arena changes the row to an open license with human votes.
Until then, cite this page when you see "#2 open WebDev" posts without screenshots of the Proprietary and AutoEval columns — the correction is part of the news.
Independent labs should publish replicated WebDev tasks with frozen harness versions. Arena movement without replication tells you preference, not your deployment ROI.
Baidu's demo portfolio on the release page also highlights game engine and CAD outputs. Those clips prove long-horizon codegen ambition, but they do not automatically transfer to WebDev Arena scores. Treat each vertical as its own evaluation budget.
If your procurement spreadsheet already lists DeepSeek or Qwen as the default open WebDev stack, keep that row until ERNIE Preview weights and human vote counts catch up. Changing vendors on AutoEval alone invites rollback risk when the official Apache drop ships with different quantization or tool schemas.
Share this post with anyone who forwards Arena screenshots without license columns. The useful correction is not "Baidu lied" — it is that open and Proprietary mean different obligations for reproducibility, pricing, and compliance.
When Apache weights land, re-run the same WebDev harness you used for DeepSeek and publish the diff. That single replication does more for the field than another week of leaderboard screenshots.
A checklist for reading any leaderboard row
The same four checks apply to every model launch, not only this one:
- License column. Does the row say Proprietary or an open license? A promise of future weights does not change today's tag.
- Vote count. A row with N/A votes rests on an automated instrument, while rows with thousands of votes reflect human preference at scale.
- Error bar. If two scores fall inside each other's interval, the order between them is not settled.
- Date of the snapshot. Leaderboards move daily, so a screenshot from one day can be wrong the next.
Apply these four checks to the ERNIE 5.5 Flash Preview row and the story is clear: a promising score, an API you can try today, and open weights that are still a promise.
Related reading
- ERNIE 5.5 Flash Base open weights
- Baidu Unlimited-OCR-2 document parsing
- Gemini Hydrogen vs Argon comparison
- DeepSeek V4.1 Flash token claims
- Loop engineering for coding agents
- What is MCP
Arena snapshot descriptions, Qianfan pricing mirrors, and Baidu release text reflect sources checked November 13, 2026.