한국어 English ✓

Korea Is Outside the License — ERNIE 5.5 Flash Base, the Free 280B That Topped 24 of 30 Benchmarks and Still Cannot Be Used in Korea

As of Nov 10, 2026 · Checked directly against Baidu's official ERNIE announcement, official model card, and the full license text · Views -

When a new open-weight base model lands, teams shopping for a fine-tuning starting point usually open the scorecard first. ERNIE 5.5 Flash Base, which Baidu released on November 5, punishes that order of reading. The first thing to read is not a benchmark but the first definition in its license: the permitted territory is the whole world except the European Union, the United Kingdom, and South Korea — and South Korea is named on that exclusion list. This article separates the scores and specs the official model card settles, the price conditions the license text settles, and the ground no one outside the company has verified yet.

The short version

What shipped is not a finished product but the end-of-pretraining checkpoint of a 280B/14B base model, and downloading it costs $0 (official announcement, checked directly)

The price hangs not on a per-token rate but on the license ladder — non-commercial use is free, commercial use at or below $50M in related annual revenue is also free with attribution and a one-time notice, and above that a prior written authorization is required (full license text, checked directly)

In South Korea, the EU, and the UK this is not a question of price at all: outside the Territory, the model, its derivatives, and their outputs may not be used (§1.11, §5.3)

ERNIE 5.5 Flash architecture diagram from the official announcement
The ERNIE 5.5 Flash architecture diagram from Baidu's official announcement: a 43-layer backbone combining hybrid sparse attention, a multi-router MoE, and ECHO. Image: Baidu ERNIE official announcement, "ERNIE 5.5 Flash Base"

Half-Open 280B: A Release That Stops at the End of Pretraining

What came out is an ingredient, not a product. A sparse MoE model of 280B total parameters with 14B active per token was pretrained from scratch on 33.5T tokens, and the checkpoint from just before mid-training and post-training began is what was published. It has not yet learned to follow instructions or hold a conversation, and Baidu itself frames the release as a common starting point for domain adaptation, instruction tuning, and reinforcement-learning research. A post-trained ERNIE 5.5 Flash is announced as following, with no date attached.

The design aims throughout at lowering the cost of further training and of running the model. Of its 43 layers, 37 use Heavily Compressed Attention, which compresses every 128 tokens into a single cache entry, while only 6 layers attend to the full context token by token. Its routers — plural — choose 10 of 512 experts, and future-token prediction is handled not by a bolt-on module but by the final layer itself, which doubles as the draft model for speculative decoding. Training context ran to 128K, and fine-tuning and inference support up to 1M tokens — but read that as a spec-sheet ceiling: both official serving examples (SGLang and vLLM) default to a context length of 131,072, with 1M listed separately as the server maximum.

The Scorecard: First in 24 of 30, With the Six Losses in the Same Table

Performance comparisons only make sense among base models. Baidu measured 30 benchmarks against the three same-tier bases public at launch — DeepSeek-V4-Flash-Base (284B/13B), Hy3-preview-Base (295B/21B), and MiMo-V2.5-Base (310B/15B) — under one harness, one set of prompts, and the same few-shot examples, and printed the full table in the official model card.

Benchmark (Baidu unified pipeline, vendor-measured)ERNIE 5.5 Flash BaseDeepSeek-V4-Flash-BaseHy3-preview-BaseMiMo-V2.5-Base
MMLU (5-shot)93.389.487.587.6
MMLU-Pro (5-shot)78.469.266.965.7
GPQA (5-shot)60.255.950.251.4
HumanEval+ (1-shot)82.369.578.074.4
LiveCodeBench (1-shot)40.131.933.532.4
RepoQA (0-shot)91.781.091.356.5
LongBench-v2 (0-shot)48.744.539.844.6
FullStackBench (1-shot)54.643.356.754.1
InfiniteBench (0-shot)46.548.033.046.7

Counted in full, ERNIE leads in 24 of the 30 events, with the widest gaps in code and knowledge — 12.8 points ahead of the DeepSeek base on HumanEval+, 9.2 points on MMLU-Pro. The six losses are in the same table: Hy3-preview takes FullStackBench, DeepSeek takes InfiniteBench, HELMET-Rerank, and ChineseSimpleQA, and INCLUDE-lite is lost narrowly, 79.5 to 79.8. The accurate reading is not "swept the board" but "broadly ahead on exams, code, and knowledge, splitting long-context and practical coding events."

Overall evaluation results chart from the official ERNIE 5.5 Flash Base announcement
The overall results chart from Baidu's official announcement. Competitors and conditions follow Baidu's own unified pipeline. Image: Baidu ERNIE official announcement, "ERNIE 5.5 Flash Base"

Two caveats attach. First, the table was designed and measured by Baidu itself, and no independent body has re-measured it; running all four models under one harness is a gesture toward fairness, but the referee is still the vendor. Second, the competitors are arguably not each vendor's newest base — DeepSeek has V4.1-Flash-1017, Tencent has Hy4, and Xiaomi has MiMo-V2.6-Flash — a point that, in this pass, stays at the level of reported commentary. And few-shot scores for a base model measure the quality of the pretrained material, not instruction following, tool use, or multi-step task completion.

The Price: $0 to Download, Priced by Territory and Revenue

There is no per-token price. Baidu publishes no hosted API rate, and the weight files themselves download for $0. The price instead hangs on the license ladder, which the full license text sets out directly.

Use (full license text, checked directly)License feeConditions
Non-commercial use (research, education, evaluation, personal experiments)$0No commercial notice or separate authorization needed
Commercial use, related annual revenue at or below $50M$0Prominent "Powered by ERNIE" attribution + a one-time notice to Baidu before first use
Commercial use, related annual revenue above $50MUnconfirmed (set by separate written agreement)Prior written authorization from Baidu required; unauthorized use above the threshold must stop
Use in South Korea, the EU, or the UKNot a pricing questionOutside the Territory — the model, derivatives, and outputs may not be used or distributed

The last row is the heaviest line in this model's price. The first three rows hold only inside the Territory. For a team in Korea, before any question of free or not, the license offers no lawful way to use the model at home at all. Routing around it through a foreign entity or server is also closed on the text: §5.3 places using the outputs outside the Territory beyond the grant. Derivative models carry the same territorial limit, and distributors remain responsible for recipients' compliance.

For teams that clear the territory gate, the real bill is hardware. "14B active" is a statement about compute, not memory: serving the model means holding all 280B of weights in memory. The roughly 288GB size of the FP8 files is, in this pass, a secondhand figure and is not asserted here — but both official serving examples defaulting to 4-way parallelism (--tp 4, --data-parallel-size 4) settles that this is not a single-GPU job. The three rival bases also compete at a $0 download price, so in this market the price comparison is not per-token rates but "in which territory, under what conditions, on how much hardware." The rivals' own license terms were not checked against their primary texts in this pass, so they are not in the table.

Where It Lands: Material for Teams That Train, Not Yet a Product

Inside the permitted territory, the official documents narrow the uses to three: continued training on your own domain data, instruction tuning that gives the model your service's voice and formats, and research into post-training methods themselves, reinforcement learning included. A fixed base checkpoint is, for experiments comparing training methods fairly, an advantage rather than a limitation: nobody repays the cost of pretraining, so each team's budget goes entirely into its own post-training.

The ecosystem moved on launch day, by current reporting: vLLM announced day-0 support, SGLang shipped a dedicated prebuilt image, and NVIDIA's Megatron Bridge gained a full fine-tuning recipe. The serving path is a single completions endpoint. As a text-continuation model it can be probed without fine-tuning on few-shot classification, extraction, and generation — but with no chat format and no tool calling, it cannot be dropped straight into a support desk or an agent seat.

Real adoption cases outside the company are, at this early stage, outside what can be confirmed. No hosted inference provider carries it yet (per reporting of the Hugging Face page), and no independent scorecard exists. The thickest external signal so far is the launch thread on Hacker News, where — again by reporting — welcome for a base-model release sits alongside the practical objection that without instruction tuning it is no coding agent, and criticism of the license's territorial carve-out. The thinness of this section says less about the model than about timing: the post-trained version and other people's re-measurements have not had time to accumulate.

Read With Care: A Self-Refereed Scorecard, a Missing Korean Row, and the Weight of a Territory Clause

First, who refereed. The 24-of-30 count is an honest count of Baidu's table, but the table itself was built with Baidu's harness and Baidu's chosen rivals. Alongside the observation that the rivals are not the newest generations, reporting also carries DeepSeek's own published MMLU-Pro figure differing from Baidu's measurement (68.3 versus 69.2) — so weight the table as "relative order under one shared setup," not as absolute values.

Second, the Korean blank. The 30 official events include Chinese exams (CMMLU, ChineseSimpleQA) and multilingual bundles (MMMLU-lite, INCLUDE-lite, MultiLoKo), but no row measured on Korean. For Korean readers the blank is double: the license bars domestic use, and even if that condition lifted, no official number yet speaks to the model's quality as Korean material.

Third, the territory clause is not a footnote to skim but the model's core condition. The territorial limit follows derivatives and outputs; the authorization procedure at the moment revenue crosses $50M — and the fee for that band, which the license does not state — is a first-order practical question. That distillation into other models is closed on the text is also no small limit in an era when training other models on synthetic data is routine.

The conclusion splits cleanly. Confirmed: a 280B/14B base checkpoint that finished 33.5T tokens of pretraining is free to download; a scorecard of first place in 24 of 30 events, measured under one shared setup, is printed in full in the official model card; and the license text excludes South Korea from the permitted territory. Not yet confirmed: whether the ranking survives independent measurement, when the post-trained version ships and under what license, the actual fee above the $50M line, and Korean-language quality. For a team in Korea, the practical reading is simple: this is not a model for today's shortlist, but one to watch for whether the post-trained version's license keeps the same territorial clause. Until that clause changes, its scorecard remains, in Korea, somebody else's scorecard.

Sources

Puzz, in one line

Judge ERNIE 5.5 Flash Base not by its scorecard but by the first clause of its license. In Korea it is a model you may download but may not use, ribbit.