Baidu's ERNIE 5.5 Flash Preview Bets on 14B Active Parameters, With Apache 2.0 Weights Still a Promise
Baidu's 280B-total, 14B-active ERNIE 5.5 Flash arrives as an open base model and a Qianfan-hosted Preview, with a 1M-token context and Apache 2.0 weights reportedly due November 25. We separate what Baidu has documented from what is still a claim.
Baidu has not been the loudest name in the Chinese open-weight race this autumn, with DeepSeek, Alibaba, Tencent and Meituan all shipping flagship models. On November 5 and again on November 11, it made its own move in the most technical way available: by publishing a base model with a full architecture write-up, then following it six days later with a post-trained checkpoint. The result, ERNIE 5.5 Flash, is a 280-billion-parameter sparse mixture-of-experts model that activates only 14 billion parameters per token. It matters less for any single benchmark than for what it shows about where Chinese labs now compete: efficiency per active parameter, long-context cost, and the speed at which a checkpoint becomes a product.
This piece separates what Baidu has actually documented from what is still a claim, a report, or a number that does not reconcile across sources.
What Baidu actually shipped
The sequence matters. On November 5, Baidu released **ERNIE-5.5-Flash-Base**↗, the raw end-of-pre-training checkpoint, with no mid-training or post-training applied. Its own description is unambiguous: a sparse MoE language model with 280B total parameters, 14B activated per token, pre-trained from scratch on 33.5T tokens, with a training context of up to 128K and support for a 1M-token context at fine-tuning and inference time. The weights are on Hugging Face in FP8 and BF16, for example baidu/ERNIE-5.5-Flash-Base-BF16↗, under a custom "ERNIE 5.5 Flash Base Model License Agreement v1.0" rather than a standard open-source license.
On November 11, the company followed with **ERNIE 5.5 Flash Preview**↗, which Baidu calls the first post-trained model built on that base. Same 280B/14B shape, same 1M-token window, now served through the Qianfan platform's web interface and API. Baidu's release page says the weights "will be open-sourced under Apache 2.0 soon"; a report relayed by the AI Radar China digest↗, citing Leiphone, puts the deadline at November 25.
Three facts deserve emphasis because they set the terms of the whole release:
- The Preview is explicitly unfinished. Baidu describes it as an intermediate post-training checkpoint and warns that some capabilities may fall short because training is not complete; an official ERNIE 5.5 Flash release is still to come.
- The Preview is text-only. In the game-development demonstration Baidu notes the model "is a text-only model and can't see the screen", so it converted playtest problems into measurable tests in code instead of looking at output.
- Open weights are a promise, not a fact. Only the Base checkpoint is downloadable today, and under a bespoke license; the Apache 2.0 release of the post-trained model is a stated intention with a reported date.
"The weights are the checkpoint at the end of pre-training, before any mid- or post-training, and can serve as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research." — Baidu ERNIE team, ERNIE 5.5 Flash Base release notes
Inside the architecture: efficiency as the design goal
Baidu's Base write-up is unusually detailed, and it is the most substantive part of the release. The model has 43 Transformer layers, a hidden size of 4096 and 64 attention heads. Nearly every component has been rethought around one question: what does it cost to sample from this model repeatedly, over very long contexts, in agentic loops?
Hybrid attention that caches very little
Only one layer in seven keeps token-level access to the full context, using Multi-head Latent Attention (MLA). The rest use what Baidu calls Heavily Compressed Attention (HCA), which compresses every 128 tokens into a single cached entry and attends over those entries plus a short window of recent tokens. In total, 37 of the 43 layers see the sequence compressed 128-fold, and the six remaining MLA layers use sparse attention (DSA) with a lightweight indexer that picks the top-k past positions per query. The practical consequence is that per-query attention cost in those layers stays constant as context grows. Baidu reports that this design still performs strongly on six long-context benchmarks.
Multi-router MoE with latent experts
The feed-forward blocks use 512 routed experts (top-10) plus one shared expert. Baidu argues that a single router's scores correlate only weakly with how useful each expert actually is, especially in middle and deeper layers, so it replaces the single router with several independently initialized sigmoid routers that score experts jointly. A second trick, "latent experts," projects tokens from 4096 to a 2048-dimensional space before the experts run, which Baidu says roughly halves dispatch/combine traffic and expert FLOPs versus full-width experts.
Smaller changes that add up
The release also lists several less obvious modifications, each aimed at training quality or inference cost:
- Manifold-Constrained Hyper-Connections replace standard residual connections with four parallel residual streams and a constrained mixing matrix, at negligible extra compute.
- ECHO replaces the conventional multi-token-prediction module by reusing the backbone's own final layer to predict future tokens, and the same layer doubles as the draft model for speculative decoding.
- MultiMax output head swaps a single softmax temperature for a learnable piecewise temperature, which Baidu says lowers pre-training loss with no extra memory.
- A new ~200K BPE tokenizer that tokenizes numbers right-to-left in groups of up to three digits, so digits of equal place value line up across numbers of different lengths.
Training used the Muon optimizer and FP8 mixed precision, plus a custom attention kernel, FlashMask-4, which Baidu says is up to 175% faster in the forward pass and 74% faster in the backward pass than FlashAttention-4 MaskMod on SM100 GPUs.
Readers who followed Neuron's coverage of Ant Group's Ling-3.2-flash↗ will notice the parallel: another 1M-context MoE in the 280-to-560B class that spends its engineering budget on activated compute rather than headline size.
The benchmark picture: strong base, thin post-training evidence
The credible numbers in this release belong to the Base model, because Baidu published a full table. It compared ERNIE-5.5-Flash-Base against three open base models of similar size under the same harness, prompts and few-shot examples: DeepSeek-V4-Flash-Base (284B total, 13B active), Hy3-preview-Base (295B/21B) and MiMo-V2.5-Base (310B/15B).
On Baidu's evaluation, the 14B-active model leads on the large majority of rows. Selected results, all from Baidu's own table:
- MMLU 93.3 against 89.4 for DeepSeek-V4-Flash-Base, 87.5 for Hy3-preview-Base and 87.6 for MiMo-V2.5-Base; MMLU-Pro 78.4 against 69.2, 66.9 and 65.7.
- GPQA 60.2 against 55.9, 50.2 and 51.4, and LiveCodeBench 40.1 against 31.9, 33.5 and 32.4.
- HumanEval+ 82.3 and MBPP+ 81.5, ahead of all three comparators on both.
- RepoQA 91.7, essentially tied with Hy3-preview-Base at 91.3 and far ahead of MiMo-V2.5-Base at 56.5.
The losses are informative too. DeepSeek's base model keeps the lead on ChineseSimpleQA (75.3 versus 73.0), HELMET-Rerank (60.0 versus 56.2) and InfiniteBench (48.0 versus 46.5), and Hy3-preview-Base edges ERNIE on BBH with chain-of-thought and on FullStackBench (56.7 versus 54.6). So the fair reading is not "Baidu beats everyone" but "Baidu is competitive to ahead on a self-run, like-for-like base-model comparison, with DeepSeek still stronger on some long-context and Chinese-knowledge tasks."
For the post-trained Preview, the picture is much thinner. Baidu's release page leads with demonstrations rather than a leaderboard, and the benchmark appendix is not reproduced in the page text I could retrieve. The AI Radar digest, citing Leiphone, says Baidu claims first place on AutomationBench and SpreadsheetBench V2. Treat that as a vendor claim until independent runs appear. This is the same caution that applied when Alibaba opened Qwen4-Max-Preview↗ with no published scores.
An efficiency story only becomes a ranking story once someone other than the vendor runs the evaluation. For a checkpoint Baidu itself calls intermediate, that has not happened yet.
What the Preview is meant to do
Baidu's demonstrations are agentic and long-horizon, and they reveal the product thesis more clearly than any table:
- Game development. In one 11-hour coding task the model merged assets from several generations of Unreal Engine into a UE 5.8 project and wrote an entire third-person wuxia game in C++, and in another it built nearly 100 mobile mini-games from design documents.
- 3D modeling and CAD. Using the Blender Python API and FreeCAD scripts, the model built an airport and airliner, a temple complex, a W16 engine and a lunar-rover concept. Baidu credits external assets such as OpenStreetMap data and Poly Haven textures.
- Knowledge work. Baidu says it built 300,000 workspaces holding 10 million reference files across more than 40 file types for finance and law training, and showed the model producing 20-page and 24-page research reports and a Monte Carlo wealth-planning workbook.
- Video creation. More than 150,000 creation environments, over a thousand creation skills and about 5,300 minutes of finished video underpin demos such as a 116-second Python-drawn animation driven by more than 10,000 lines of code.
These are curated showcases, not controlled tests, and Baidu says as much by crediting external tools alongside each piece. They are still a useful signal of training emphasis: the Preview was optimized for tool use inside an agent harness, not for chat.
Pricing, and a number that does not reconcile
Baidu's release page says the API is live on Qianfan but, in the text retrieved, does not state prices. The pricing that circulates comes from the AI Radar digest citing Leiphone and TMTPost: 2 yuan per million input tokens at peak and 1 yuan off-peak, with output at 8 yuan peak and 4 yuan off-peak, plus a limited-time 50% discount. If accurate, that is cheap for a 1M-context model, and it mirrors the peak/off-peak structure that DeepSeek introduced earlier this year↗.
The same digest, however, reports the model as having "2800 billion total parameters and 140 billion active" — ten times the figure on Baidu's own release pages. That is almost certainly a transcription error, since the Base and Preview pages both say 280B and 14B, but it is a reminder of why the primary source should win when numbers conflict. Readers should confirm current Qianfan rates directly before budgeting.
The competitive field: three flagship launches in eight days
Baidu's timing is not accidental. Tencent released Hy4 on November 4↗, describing it as an open-source model and reporting a 3.04 score on a financial-task evaluation of 579 real tasks, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68. Meituan's LongCat-2.5 open-source announcement↗ describes a 1.6-trillion-parameter MoE with about 48B active parameters, a 1M-token context and an MIT license permitting commercial deployment and further training.
Set side by side, the three illustrate different bets:
- Meituan pairs the largest total size with the most permissive license, shipping weights under MIT.
- Tencent is emphasizing enterprise agent reliability and computer use, with open-source positioning but no named license in the announcement text.
- Baidu is the smallest by a wide margin, 14B active against roughly 48B for LongCat-2.5, and is betting that a lean active footprint plus aggressive cache compression wins on cost, while still owing the community its Apache 2.0 weights.
The license point is the sharpest one. Meituan's MIT weights are downloadable now; Baidu's post-trained weights are not, and its Base license is custom. For developers choosing a model to fine-tune this month, that gap is practical, not rhetorical.
What to watch next
Four checkpoints will determine whether ERNIE 5.5 Flash becomes a reference model or a footnote:
- November 25. The reported Apache 2.0 weight release for the post-trained model. A slip would hand the open-weight narrative to Meituan and Tencent.
- The official ERNIE 5.5 Flash release. Baidu says post-training is ongoing and the Preview is improving quickly, so benchmark numbers for the final version should be treated as the real test.
- Independent evaluations. Third-party runs on agentic and coding benchmarks will show whether the efficiency claims hold outside Baidu's harness.
- Reproducing the architecture. Because the Base weights and design notes are public, researchers can test claims such as multi-router MoE and 128x compressed attention directly. Any confirmed gain would travel quickly to other labs.
There is also an infrastructure subtext. Baidu says its automated research system, ERNIE Evolve, was used for inference optimization on this model, reporting generation speed rising to 623.6% of baseline in low-latency scenarios on its Kunlun P800 chip, according to the AI Radar summary of Baidu's technical blog. If that holds in production, it is a notable data point for Chinese labs optimizing on domestic silicon.
Bottom line
ERNIE 5.5 Flash is the clearest signal in months that Baidu intends to compete on open, efficient, long-context models instead of conceding that ground. The 280B/14B design, the 33.5T-token pre-training run, and a base-model table that leads most rows against comparable open models are real and verifiable on Baidu's own pages. What is not yet verifiable is how the post-trained Preview ranks, what the final weights license will look like, and whether the November 25 date holds.
For practitioners, the advice is simple: pull the Base checkpoint if you do research or domain adaptation and can live with the custom license; test the Preview through Qianfan for agentic coding and office-automation workloads; and wait for independent numbers and the Apache 2.0 weights before committing production traffic. The other Chinese labs will not wait.
Links & Resources
External links — opens in a new tab

The TI-Nspire CX II CAS Treatise
by Richard Murdoch Montgomery
A comprehensive guide covering CAS programming, 3D graphing, calculus, linear algebra, and physics applications on the TI-Nspire.

Calculus I
by Richard Murdoch Montgomery
Limits, derivatives, integrals, and series — a first course in calculus with formal proofs, worked examples, and applications to physics and engineering.

Machine Learning in Forensic Anthropology
by Richard Murdoch Montgomery
Applying SVMs, CNNs, and ensemble methods to skeletal identification, age estimation, and ancestry determination in medico-legal contexts.

Partial Differential Equations: Theory, Methods, and Applications
by Richard Murdoch Montgomery
A rigorous, modern treatment of the heat, wave and Laplace equations — the math that underpins the physics of computation.
Comments
Open discussion — no account needed. Be respectful.
More from Chinese Models Desk
Meituan Gives Away LongCat-2.5, a 1.6T-Parameter Agent Model, Under MIT Terms
Meituan open-sourced LongCat-2.5, a 1.6-trillion-parameter multimodal MoE trained on domestic Chinese chips, under the MIT license with no revenue thresholds. The food-delivery giant is betting that permissive weights, not a token business, buy it a place in the developer toolchain.
DeepSeek Ships Its Whole App as an API: Search, Files, Voice and Mandatory Safety Checks on a $0.30 Flash Model
On November 13, DeepSeek exposed the features behind its consumer app through its API: server-side web search, file parsing, URL reading, OCR, ASR and TTS, all anchored on deepseek-flash. The price list and the compliance clause are the real story.
Alibaba Opens Qwen4-Max Preview: 1M Tokens, Two Weeks, and No Benchmarks
Alibaba's Qwen4-Max preview offers a 1M-token window and video input through Token Plan until November 25, but ships without benchmarks, pricing, weights or a license.