Deep Dive — ERNIE 5.5 Flash: 280B params for ¥2 a million tokens

Baidu released ERNIE 5.5 Flash Preview this morning: 280 billion total parameters, 14 billion of them active per token, a 1-million-token context, and a Chinese list price of ¥2 per million input tokens at peak hours — roughly $0.28. It is the first post-trained version of the base model Baidu published on Hugging Face six days ago, and it is live today through Qianfan, Baidu's cloud, while the weights stay locked until November 25.

So the honest answer to the question you are probably typing — is this worth switching to? — is: only if you need a 14B-active model with a million-token window this fortnight and can live with a preview checkpoint. Baidu says as much on its own release page: the model is "an intermediate checkpoint from post-training," already competitive with models of its size, but "because training is not finished, some capabilities may still fall short," according to the ERNIE 5.5 Flash Preview release.

What makes the release worth 2,000 words is not the spec sheet. It is the second half of the announcement, which Baidu buried nine days earlier in a research post: the production inference stack for this model was partly tuned by an automated research system that Baidu built, ran for a week at a time, and is now describing as its first step toward recursive self-improvement.

The release, in numbers

The architecture is the same one Baidu published on November 5 with the base model: 43 Transformer layers, a hidden size of 4096, 64 attention heads, a sparse mixture-of-experts backbone with a multi-router latent MoE, and a training run of 33.5 trillion tokens at up to a 128,000-token context. Fine-tuning and inference stretch that window to 1 million tokens. The base checkpoint shipped in both FP8 and BF16 on Hugging Face, per Baidu's base-model release; The Decoder noted at the time that Baidu benchmarks it against DeepSeek V4 Flash and MiMo V2.5 — a full 280B checkpoint is around 280 GB in FP8, so this is a multi-GPU deployment, not something you run on a workstation.

The interesting structural choice is attention. Only one layer in seven keeps token-level access to the entire context, using multi-head latent attention; the rest use what Baidu calls heavily compressed attention, which collapses every 128 tokens into a single cached entry and then attends over those entries plus a short recent window.

That is a 128-to-1 compression of the KV cache in most of the stack, which is how a 1-million-token window stays servable at all. The token-level layers use sparse attention with a lightweight indexer that scores past positions and computes real attention only over the top-k, so their cost per query stays flat as the context grows. Cheap layers get wide 512-dimension heads, expensive layers get narrower 256-dimension ones — the opposite of the uniform layout most models use.

The trade-off is real and Baidu does not hide it: most layers see the distant past at 128-token granularity. That is fine for repository-scale code, long documents and multi-turn agent logs, where you need gist and consistency rather than exact recall of token 412,000. It is not the setup you want for byte-exact retrieval from the middle of an 800K-token file, and no independent long-context evaluation of this checkpoint exists yet. We have written before about how long context costs 3.6x to 7x more per token than short context even with compression tricks like this; the architecture lowers the multiplier, it does not repeal it.

What it costs, exactly

Baidu's Qianfan price list, as reported by Leiphone, is ¥2 per million input tokens and ¥8 per million output tokens during peak hours, dropping to ¥1 and ¥4 off-peak, with a limited-time 50 percent discount running alongside — if that applies to the listed price, peak input lands at ¥1 per million, per Leiphone's write-up of the launch. That same piece dates the open-weight drop at November 25, which matches the Chinese release notes; the English page only says "soon."

Run the arithmetic on the list price. Pushing a 200,000-token repository through the model for an agent turn costs ¥0.40 in input — about 5.6 cents. Sending a full million-token document costs ¥2, or 28 cents, and caching makes repeat passes cheaper. For comparison, Zhipu's GLM-5.3-Flash was priced at roughly $0.045 per task when we worked through its economics in August, so the two labs are converging on the same order of magnitude from different directions: Baidu gets there with a bigger model and a longer window, Zhipu with a leaner one.

Developers outside China can already reach it. OpenRouter lists baidu/ernie-5.5-flash-preview served by Qianfan at $0.28 per million input tokens and $1.13 per million output, with cache reads at $0.028 per million, per OpenRouter's model listing. Those dollar figures are the list price with no promotional discount — the 50 percent cut appears to be a China-only channel offer.

The timing is not accidental. QbitAI reported today that Zhipu, MiniMax, Alibaba Cloud and Qoder all put model subscription cards on Tmall at discounts of 50 percent or more for Singles' Day, the first year tokens have been sold as a standalone consumer category on the platform. Chinese model pricing is now set by a retail calendar as much as by compute costs, and a launch-day discount is table stakes.

The model that helped tune itself

Nine days before this launch, Baidu's ERNIE team published the ERNIE Evolve research post — an automated research system given a code repository, an optimization objective and an evaluation harness. It works in two nested loops: an outer loop picks which research direction to pursue next, an inner loop modifies code and submits each candidate for independent evaluation and review. Five roles split the work — Director, Selector, Executor, Evaluator, Reviewer — and only changes that pass evaluation and review are kept, according to Baidu's ERNIE Evolve writeup.

The public-benchmark results are the cleanest part of the claim, because they are measurable against published numbers. On NanoChat AutoResearch it reached 0.8763 validation bits-per-byte against a best public result of 0.8878. On NanoGPT Speedrun it hit the target loss in 38.9 seconds versus 39.9 seconds, and in 2,590 steps versus 2,690. On NVIDIA's SOL-ExecBench kernel-optimization task it scored 0.819 mean SOL — ahead of the two other automated research systems Baidu compared against, but behind the 0.827 best public result.

Then the part that actually shipped. Baidu pointed the same loop at the inference stack of the model in development, ERNIE 5.5 Flash, and reports three campaigns. On Kunlun P800, its own accelerator, in a low-latency setting, the system ran about 168 hours across four research directions and lifted generation speed to 623.6 percent of the starting code. On deterministic inference throughput it spent three days across three directions to reach 133.7 percent, cutting per-token time 26 percent and first-token latency 12 percent. On decoding speed it explored seven directions, landed three, and ended 14.9 percent faster than baseline at unchanged accuracy.

Read those three numbers against each other and the story gets more interesting than the headline. A 6.2x speedup after four directions of search is not a model discovering something a human team missed; it is a search loop finding that the starting point was hand-written and unoptimized. Baidu's own post says the largest single gain in the decoding campaign came from replacing a sequential expert scan in the MoE scheduler with a two-level index, and that another chunk came from noticing that FP8 matrix multiplies on inputs under eight rows were half zero-padding. Those are the wins a good performance engineer finds in a week — the system's contribution is that it found them unattended, tested them, and kept only the ones that passed review.

Which is the fair way to read "recursive self-improvement" here: an autotuner with a strict evaluator and a knowledge base, improving kernels and schedules, not a model improving its own weights. The 1 to 3 percent edges on public benchmarks are real but inside the noise band of baselines that hundreds of researchers have already squeezed. Baidu's own framing is more careful than the phrase suggests — it calls inference infrastructure "the first landing point" for RSI and commits to keeping the loop interruptible and under human oversight.

Why give away the base and sell the finished model

Baidu's commercial position explains the release cadence better than any technical consideration. The company lost the consumer chatbot war to ByteDance's Doubao and Alibaba's Qwen, and its growth now comes from the layers underneath: cloud, chips, model APIs. In the second quarter, AI businesses were half of Baidu's core revenue and GPU cloud grew 283 percent, even as search advertising fell 19 percent, as we covered in the Q2 numbers.

Free base weights serve that strategy precisely. Publishing the pre-training checkpoint commoditizes the most expensive stage of the pipeline — 33.5 trillion tokens of pretraining — and pushes the value into post-training, serving and the chips underneath, which is where Baidu actually owns assets. The twenty-day gap between the API launch and the weight release is the monetization window: anyone who needs ERNIE-class output this month pays Qianfan, and anyone who wants to fine-tune waits.

The contrarian read is that this is a company with fewer options than it looks. Apache 2.0 on the post-trained weights means Qianfan's advantage is price and Kunlun capacity, not the model itself — a commodity business with a heavy balance sheet. The labs that hold weights back (Anthropic, OpenAI, Moonshot on its frontier tier) are the ones collecting rents; the ones giving them away are buying distribution and hoping the ecosystem returns the favour. Baidu has been on both sides of that line this year, and the November 25 date tells you which side it ended up on.

What could go wrong

The most obvious risk is that the preview is genuinely a preview. Baidu publishes its benchmark results as an image, not a table of numbers in text, and its English post says only that the model scores "competitively" while Leiphone names Terminal-Bench and AutomationBench as the strong suits. That is a company choosing not to put a number in a sentence, which is worth noticing.

The second is context economics under real traffic. A million-token window priced at ¥2 of input assumes customers send prefill-heavy, cache-friendly requests. Agent loops that rewrite the context every turn, with 300K-token working sets, will bill very differently — and the token-level MLA layers still keep every token in cache, so serving cost scales with the number of concurrent long sessions, not with the compression ratio.

The third is silicon. The 623.6 percent figure is a Kunlun P800 number, and Baidu's ability to serve this model cheaply depends on how many P800s it can deploy. Baidu does not publish that number, which means the price is credible but the capacity behind it is unverified.

What to watch

November 25 is the date that matters. If the weights land under Apache 2.0 as the finished model rather than the preview checkpoint, ERNIE becomes the largest permissively licensed Chinese model with a genuine million-token window, and the fine-tuning ecosystem gets a new base overnight — the vLLM and SGLang paths already exist for the base model, and NVIDIA's Megatron-Bridge has an ERNIE 5.5 fine-tuning example.

After that, three measurements decide whether this was a launch or a moment. Independent benchmark runs — Artificial Analysis or LMArena placement, and anyone rerunning Terminal-Bench on the API — will tell you whether "competitive" means within a point of DeepSeek V4 Flash and MiMo V2.5, the models Baidu benchmarks against, or a tier below. Whether the 50 percent discount survives December will tell you whether this is a price war or a promotion. And whether ERNIE Evolve's loop graduates from inference kernels to training-system optimization, which Baidu says is next, will tell you whether the RSI story has a second act or was a launch-week footnote.

For now: a 280B open-weight-class model at 28 cents a million input tokens, with a million-token window and a self-tuning stack behind it, is the cheapest way to find out how much of your workload is actually long-context.

Would you trust an agent-tuned inference stack more or less than a hand-tuned one? Tell us in the comments.