Baidu used its annual Baidu World conference on 12 November to open Token Exchange, a marketplace where businesses compare and buy AI model services from multiple providers, and to put its newest model family behind it: ERNIE 5.5 Flash Preview, released 11 November, positioned as the workhorse for agent workloads. Per Unite.AI’s coverage of the launch, Baidu claims the marketplace is already averaging more than 50 million calls and over 700 billion tokens a day after a month of internal operation, with more than 100 model resources onboarded.
The marketplace idea is straightforwardly aimed at a real market failure. Token Exchange verifies providers by comparing model fingerprints and running capability, performance and availability tests — Baidu describes a system of credit ratings built on regular spot checks and real-time end-to-end call monitoring rather than a one-time listing check. Anyone who has bought inference through a reseller will recognise the problem being solved: you are buying a claim about which model is actually behind the endpoint, and until now the only verification was vibes. Fingerprinting plus continuing oversight is Baidu asserting that its own marketplace vouches for what it sells.
The model underneath it is the more consequential half. ERNIE 5.5 Flash Base was released 11 November as a sparse Mixture-of-Experts language model with 280 billion total parameters and 14 billion activated per token, pre-trained from scratch on 33.5 trillion tokens with a training context of up to 128K and support for a 1-million-token context at fine-tuning and inference — per Baidu’s own release notes. The base weights are live on Hugging Face, though under a Baidu custom licence rather than a fully open one. The post-trained Preview Baidu positions at the front of the stack — priced through its Qianfan platform at RMB 1 per million input tokens and RMB 4 per million output — is promised under Apache 2.0 by 25 November, which would put a frontier-adjacent Chinese model under one of the most permissive licences in use.
Baidu’s demonstration of the Preview is telling about where Chinese labs think value now sits: in an 11-hour Unreal Engine task, the text-only model wrote a third-person game in C++ and converted playtest bugs — a hovering character, a misplaced camera — into measurable software tests. It could not see the screen; it worked entirely from descriptions and build logs. Agent-harness work, not chat, is the frame.
What this changes
Three patterns worth naming.
First, the pricing floor keeps falling, and it is falling fastest on exactly the model class agents consume. RMB 1 per million input tokens is roughly a fifth of a US dollar at current rates; OpenRouter already lists the Preview’s free tier. Combined with Anthropic’s 20% price cut on Fable 5.2 earlier this month, as VentureBeat reported at the launch, the agentic-token market is in a commodity squeeze, and every lab’s margin is now set by whoever has the cheapest compute, not the cleverest model.
Second, the licence structure is deliberate. Base weights under a Baidu-controlled licence, post-trained weights under Apache 2.0, is the inverse of the usual Western pattern (Apache base, safety-clause fine-tune). Releasing the post-trained checkpoint permissively while gating the base is aimed squarely at the same developers Alibaba courted with Qwen’s open-weight releases — and it keeps Baidu’s training recipe, the genuinely hard part, behind the custom licence.
Third, Baidu World bundled the model news with its own silicon: the Kunlunxin M100 inference chip shipping now and the M300 flagship expected 2027, plus a 1,024-card Tianchi supernode. The pitch is a full stack — chips, cloud, models, marketplace — with no Western component anywhere in the loop. As we argued when Alibaba opened its model family, China’s laboratories are converging on open weights plus domestic hardware as a system policy, and each new entrant makes the stack cheaper for anyone outside US export-control jurisdiction — including, as it happens, New Zealand integrators, who can run the Preview’s weights on whatever hardware they like the moment 25 November arrives.
The caveats: the 700-billion-token-per-day figure and the marketplace’s “100+ resources” are company-reported, with no independent audit of what the credit ratings actually measure. Whether fingerprint verification can really police bait-and-switch listings at scale is unproven — the design looks sound, but the marketplace is one month old. And the Apache 2.0 date is a promise, not a shipped artefact; Baidu’s own release notes caution the Preview is an intermediate checkpoint that “may still fall short in some capabilities.”
If 25 November holds, though, the record for the most capable fully-open post-trained model changes hands again, and it changes hands eastward.
Sources: Unite.AI (12 Nov 2026); Baidu ERNIE release notes (11 Nov 2026); Hugging Face model card; OpenRouter listing.
