← Back to blog
ERNIE 5.5 Flash Preview: Official Specs, Access, and Limits

ERNIE 5.5 Flash Preview: Official Specs, Access, and Limits

Olivia Hart

By Olivia Hart

MyClaw Editorial

MyClaw

Run Best-in-Class AI Agents Now

Run OpenClaw or Hermes Agent on managed MyClaw hosting, so your AI agent stays online, updated, and ready for real work.

AI Takeaway

  • What is ERNIE 5.5 Flash Preview? ERNIE 5.5 Flash Preview (ernie-5.5-flash-preview), the first post-trained Flash Base model (280B / 14B). It is not a final Flash release or ERNIE 5.1, 5.0, or 4.5.
  • How long does the preview last? Posts dated 11 November 2026 call it an intermediate checkpoint, with a Flash release planned. No API retirement date was published as of 13 November 2026.
  • What official specs and limits matter? Docs list a 1M context and 1M output cap, including thinking (64K default). It is text-only, thinking forced on at high. Tools work; built-in web search does not.
  • How do you access it, and what is official pricing? Access is through Baidu Qianfan, with China pay-as-you-go in yuan and a 50% promotion through 30 November 2026, and international pricing in USD.
  • Where does MyClaw fit? Use MyClaw to plan the prompt, context, and agent workflow around this preview. It does not claim the model runs there; confirm Qianfan access for this ID.

What ERNIE 5.5 Flash Preview Is — and What It Is Not

ERNIE 5.5 Flash Preview is the first post-trained model built on Flash Base. Baidu announced it on 11 November 2026 and opened the API on Qianfan the same day. The model ID is ernie-5.5-flash-preview. Some international rows label the model "ERNIE 5.5 Flash" while still using this Preview ID.

The 11 November 2026 posts call it an intermediate checkpoint, not the later official Flash release. Training is not finished, so some capabilities may still fall short, and improvements are planned for that release. As of 13 November 2026, no API retirement date has been published. Do not treat the open-weight plan or the China promotion deadline as a sunset for this ID. The English post says the weights will be open-sourced under Apache 2.0 soon; the Chinese post says before 25 November.

Baidu opens API access to ERNIE 5.5 Flash Preview · TechNode

Vendor copy positions it for coding, 3D modeling, knowledge work, and content creation.

This is not Flash Base, the 5 November 2026 pre-training checkpoint under a different license, and not a published final Flash ID. No non-Flash "ERNIE 5.5" product or separate Think SKU appears in the lists reviewed here. It is also not ERNIE 5.1, 5.0, or 4.5. An earlier landscape post, Best Chinese AI Model in 2026, mentions Ernie as a consumer chatbot; that is not this Preview API.

Official Specs at a Glance

The specifications below are as of 13 November 2026. This is a capability snapshot, not a quality ranking.

ItemValue
Official name / IDERNIE 5.5 Flash Preview; ernie-5.5-flash-preview
StatusPreview announced 11 November 2026; later Flash release planned; no published API retirement date
Size280B total parameters, 14B active
ModalityText in, text out
Context / max input / max output1M / 1M / 1M tokens (output includes thinking)
Default max_tokens64K (range 1–1M)
ThinkingForced on; default reasoning_effort is high; cannot disable or adjust; thinking_budget not supported
ToolsFunction calling supported; built-in web search not supported
Prompt cacheSupported
Default rate limitsChina Qianfan: 300 RPM / 3,000,000 TPM. International Qianfan: 60 RPM / 1,000,000 TPM
China pay-as-you-go (promotional / list, yuan per million tokens)Peak 8:00 to 22:00: input 1 / 2, cache 0.1 / 0.2, output 4 / 8. Off-peak 22:00 to next day 8:00: input 0.5 / 1, cache 0.05 / 0.1, output 2 / 4. Promotion printed through 2026-11-30 23:59:59; timezone not labeled
International pay-as-you-go (USD per million tokens)Input 0.28, cache hit 0.028, output 1.13. No peak/off-peak split or promotional pair in the international schedule reviewed

The 64K figure is the default max_tokens, not the 1M cap, and thinking tokens count against that shared cap. A tight default can therefore cut a long job short. The model is text only; game, CAD, or video demos used tools and credited external assets. Product copy claims 200-plus tokens per second from FP4 expert weights and a draft head; that is a vendor claim.

As of 13 November 2026, vendor-reported results appear in the 11 November 2026 announcement appendix; this article does not reprint or rank them.

How Official Access and Pricing Work

baidu/ERNIE-5.5-Flash-Base · Hugging Face

Access is through Baidu Qianfan, on an OpenAI-compatible Chat Completions API. The documented base URL is https://qianfan.baidubce.com/v2. Authenticate with a Bearer API key and set the model ID to ernie-5.5-flash-preview.

Treat the China-site and international documentation as separate schedules. China prices are in yuan, international prices are in US dollars, and default rate limits differ. The 50% promotion appears on the China product page only. Do not convert the currencies, and do not assume that promotion applies internationally.

China promotional and list prices are in the specs table. The peak and off-peak windows, and the promotion end printed as 2026-11-30 23:59:59, have no labeled timezone. The billing page lists both prices but does not restate that deadline. After the printed deadline, the list prices are the conservative figures unless a later page says otherwise.

International pay-as-you-go, from the international price page dated 10 November 2026 and fetched on 13 November 2026, is 0.28 US dollars per million input tokens, 0.028 dollars per million cache-hit tokens, and 1.13 dollars per million output tokens, with no peak/off-peak split or promotional pair.

Thinking tokens are counted in the output budget; there is no separate thinking tariff. Prompt cache is supported, and the specs table uses the model-specific China and international cache-hit unit prices. A general Qianfan prompt-cache page also states that cached tokens are billed at 40% of the prompt unit price, which does not match those unit prices. This article follows the model-specific schedules and does not try to reconcile the two.

A free quota or Token Plan inclusion for this ID was not verified and is not claimed here.

Skip the Setup, Run Best-in-Class AI Agents Now

Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.

Get Started

Practical Uses and Official Limits

Vendor copy points this preview at long-horizon coding and game-engine work, tool-using 3D or CAD modeling, knowledge work over long document sets, and agentic content workflows. Use that to judge whether the documented feature set matches the job, not as a performance proof.

The limits matter more than the feature list. The model is text-only: it cannot see a screen or a native video frame. Thinking stays on at high, so output cost and the shared 1M cap include the chain of thought, and the 64K default can cut a long job short unless you raise max_tokens. Built-in web search is not available, so live lookup has to come from your own tools. China and international rate limits and prices also differ, as shown in the table.

As noted above, this is still an unfinished checkpoint, with no published API retirement date or final Flash ID as of 13 November 2026. Do not treat ernie-5.5-flash-preview as a durable production dependency.

Preview weights were not confirmed as public in the sources reviewed here, so this is not a current self-host option. Flash Base is already downloadable under a different license, and is not this post-trained API. The English and Chinese open-weight wordings remain as stated above.

For another dated preview explainer on specs, access, and limits, see Qwen4 Max Preview.

Plan the Work in MyClaw

MyClaw is a hosted AI team that works with your tools around the clock. Use it to plan the prompt, context pack, and agent workflow around this preview, then refine that plan as work rather than as a one-off chat demo.

Start with the job, not with a model picker. Spell out the repo, document set, or tool-using 3D or content task, and size the brief to the documented 1M context, forced thinking, and shared output budget.

Get Started
  1. Open the MyClaw homepage and choose Get Started.
  2. Write the actual brief — a coding task, a long-document pass, or a tool-using 3D or content job — so the plan matches those documented limits.
  3. Use MyClaw to refine and run the surrounding workflow.

If the job requires this exact preview model, confirm access on Qianfan or ERNIE surfaces and check the current model list.

Conclusion

ERNIE 5.5 Flash Preview is a dated checkpoint announced on 11 November 2026, with documented specs, Qianfan access, and separate China and international price schedules. It is not the later Flash release. Use Qianfan if you need this exact model, and do not plan as if a final Flash ID already exists, or treat the China promotion deadline or the open-weight plan as an API sunset. Use MyClaw to plan and run the surrounding workflow, and recheck the current Qianfan model list before you depend on this ID.

Skip the Setup, Run Best-in-Class AI Agents Now

Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.

ERNIE 5.5 Flash Preview: Official Specs, Access, and Limits | MyClaw.ai