ERNIE 5.5 Flash Preview: 280B parameters, a 14B bill, and weights that arrive later

Baidu's annual Baidu World keynote in Beijing on November 12 announced a bundle: a marketplace for other companies' models, a supernode tying 1,024 accelerators together over optical links, a chip roadmap update, and four refreshed agents. Buried in that pile is the piece developers can use now. ERNIE 5.5 Flash Preview is a sparse model with 280 billion total parameters that activates 14 billion per token, shipping with a million-token context window and a price that undercuts the Western flash tier by an order of magnitude.

The model exists; the weights are a calendar entry

The API is live on Baidu AI Cloud's Qianfan platform, and the model is listed free on OpenRouter, where its page already shows traffic from coding agents such as Kilo Code. The open-weights half is a promise: Baidu says the weights "will be open-sourced by November 25," a date thirteen days out with no repository, licence or file layout attached. Until that directory exists, 280 billion parameters is a description rather than an artifact. Baidu's own headline bullet calls ERNIE 5.5 Flash Preview "the latest open-weight foundation model," which is the tense problem in one phrase.

Two benchmarks, zero scores

Baidu cites Terminal-Bench and AutomationBench by name and stops there: no numbers, no harness version, no named comparison model. The release says the model "achieved competitive scores" in the same sentence as "extremely fast decoding speed," which costs nothing to claim and nothing to check.

One of those claims can be measured indirectly, and the measurement is unkind. The OpenRouter listing describes generation above 200 tokens per second. The same page records a median throughput of 96 tokens per second across the previous three days, and a 22.75% structured-output error rate on the Qianfan endpoint. Those figures come from production traffic, not a benchmark harness, so they are no verdict on the model. They are the only independent measurements in public, and they look nothing like the marketing line.

The price is the number that holds still

Pricing is the part Baidu stated with precision, and cached input at a twenty-fifth of fresh input rewards exactly the agent workloads it keeps naming.

Unit	Baidu list price (RMB)	Converted at 6.67 yuan per dollar
Input, per million tokens	1	about $0.15
Output, per million tokens	4	about $0.60
Cached input, per million tokens	0.04	about $0.006

The conversion uses 6.6733 yuan per dollar, the rate Trading Economics recorded on November 12. Baidu bills in yuan, so the dollar column drifts with the exchange rate while the yuan column does not. For scale, Gemini 3.8 Flash lists at $1.50 input and $7.50 output per million tokens on OpenRouter, currently discounted 50%. Against that list rate, Baidu's input is exactly one tenth and its output about a twelfth. Cheap tokens are not a frontier model, and giving this one away on OpenRouter suggests Baidu knows it.

A tenth of the compute, all of the storage

The 280/14 split is ordinary mixture-of-experts arithmetic with two consequences that point in opposite directions. Fourteen billion active parameters put inference cost in the range of a mid-size dense model, which is how Baidu can charge fifteen cents. Two hundred eighty billion total parameters still have to live somewhere. At 8-bit precision, a level the release never states, that is roughly 280 GB of weights before any KV cache for a million-token window. Last week a hobbyist pushed the same trade to its limit by fitting a 320B model with 18B active into forty mini PCs chained over USB and getting 14 tokens per second. That rack is not a product, but the same memory wall will decide who can self-host the November 25 file and who can only rent it.

An RSI label and a marketplace Baidu also sells on

Baidu says the model is "optimized with ERNIE Evolve, Baidu's proprietary recursive self-improvement (RSI) technology, to boost inference efficiency and reduce cost." That sentence is the entire explanation. There is no mechanism, ablation or statement of what the model improves about itself: architecture search, kernel selection, quantisation policy, post-training data. Recursive self-improvement is contested elsewhere in the industry, and Baidu has not said what it means here. Treat it as a product label until a technical report arrives.

The marketplace is the more substantive launch. Token Exchange is described as "a neutral, transparent, and trusted marketplace connecting model service providers to customers," with verification built on model fingerprints and capability testing, and credit ratings assembled from spot checks and end-to-end call monitoring. One month in, Baidu says the exchange has onboarded more than 100 model resources and handles 50 million calls and 700 billion tokens a day. What the release does not disclose is the take rate, whether ERNIE tokens will themselves be rated on the exchange, or who audits the ratings. A marketplace run by one of its largest sellers carries an obvious conflict.

A supernode, a chip, and a story about offices

The infrastructure half followed Baidu's usual pattern: the Tianchi 1024-card Supernode, an all-optical interconnect node built for fast data movement, plus a Kunlunxin roadmap in which the M100 inference chip has shipped and the flagship M300 is expected in 2027. "We believe inference will be the primary driver of sustained long-term growth in the AI chip market," Li said.

Then the organizational thesis, where the keynote spent its energy. "If AI is the new electricity, then the intelligent organization is the new factory of the AI era, where productivity truly happens," Li said. Jobs, in his telling, converge into two types: "Builders, who create new things, and Sellers, who connect with people to forge emotional bonds and earn their trust." Tacit knowledge becomes the moat because "this is a capability that competitors can't copy, hire away, or buy." Process design gets a formula: measurability plus looping equals continuous evolution.

The evidence for all of this is internal. Average daily active agents inside Baidu rose from 5,000 to 37,000 over the past year, agents now complete more than a million tasks a week against 84,000 before, and Baidu has built 4,300 workspaces holding more than 20,000 captured "organizational skills." Those are adoption numbers for Baidu's own workforce, not market share, and the release gives no external customer counts for the four agents it upgraded.

What November 25 settles

Two things are checkable today: the API, which answers, and the price, which sits a tenth below a Western flash tier at the current exchange rate. Everything else, from the Terminal-Bench line to the RSI label to the trust score on the marketplace, is Baidu grading its own homework.

The weights are the real test. If the release lands on November 25 with a permissive licence and a file layout that fits on rented hardware, ERNIE 5.5 Flash Preview becomes a serious option for teams priced out of American labs. If the date slips, the keynote's most concrete promise will have been the one about a download that never appeared.