Baidu opens ERNIE 5.5 Flash Base weights: a 280B MoE base model with two gates—deployment and licensing
According to a MarkTechPost report dated November 5, 2026, Baidu released and opened the FP8 and BF16 weights of ERNIE 5.5 Flash Base: 280B total parameters, 14B activated per token, a 128K training context that can extend to 1M at inference. It is a base model that stops at the end of pretraining, with no mid-training or post-training; the report says the license uses a $50 million annual revenue threshold and excludes the EU, UK and South Korea.
What changed
The report says Baidu posted ERNIE 5.5 Flash Base weights on Hugging Face and ModelScope: a text-only sparse MoE base model with 43 layers and a hidden dimension of 4096, 280B total parameters and 14B activated per token, pretrained from scratch on 33.5T tokens with the Muon optimizer, with weights frozen at the end of pretraining. Baidu positions it as a unified starting point for directly comparing fine-tuning, instruction tuning and reinforcement learning methods. Architecturally, 37 of the 43 layers use heavy compression attention (HCA), caching one entry per 128 tokens, while the remaining 6 layers use MLA plus sparse attention—so 1 in every 7 layers keeps full context access. Each MoE layer has 512 routed experts (top-10) plus one shared expert, and multiple independent sigmoid routers replace a single router; ECHO reuses the backbone's last layer to predict future tokens and doubles as a built-in draft model for speculative decoding, with its extra 0.36B parameters excluded from the main parameter count. Baidu self-reports best results on 24 of 30 base-model benchmarks under a unified pipeline, with MMLU-Pro at 78.4 (DeepSeek-V4-Flash-Base at 69.2 under the same pipeline), but HELMET-Rerank at 56.2 trails that model's 60.0, and it gives up first place on 6 other benchmarks including ChineseSimpleQA, INCLUDE-lite and InfiniteBench. On deployment, the reported reference configuration is SGLang with 4-way parallelism, while vLLM requires version 0.32.0 plus the vLLM-ERNIE plugin; both accept up to 1,048,576 tokens. As a base model, it runs through the completions endpoint rather than chat.
What it means for developers
There are several direct implications for developers. First, these are open weights, not an open-source license in the usual sense: the weights are downloadable, but the license is a custom ERNIE 5.5 Flash Base License v1.0, which the report says is free below $50 million in annual revenue while excluding the EU, UK and South Korea—actual rights must be checked against the license text, and the weights being hosted on Hugging Face does not make them a standard open-source license. Second, it addresses "where to start post-training," not "plug in a chat model directly": without post-training there is no instruction following or dialogue format, so putting it into a product means either running your own SFT or RL pipeline or switching to an already aligned model, and the engineering effort differs greatly between the two. Third, the long-context cost structure is deliberately thin: HCA caches only one entry per 128 tokens across 37 layers, which should in theory cut KV cache substantially, but only 1 in every 7 layers retains full context, so how much benefit long-document retrieval and repository-level understanding actually get needs validation on your own document sets rather than taking self-reported figures like RepoQA 91.7 at face value. Fourth, the report gives only a 4-way parallelism reference configuration and does not disclose GPU models or memory, so capacity questions such as whether a single machine can serve it or how many cards are needed cannot be answered from this article. Also note that ECHO's 0.36B parameters are excluded from the main parameter count, so estimating hardware from total parameters will be somewhat optimistic.
Integration and evaluation
The following is conditional analysis or advice, not reported fact. On precision versions: the report says both FP8 and BF16 weights are released, and this first affects the subsequent training path. If you only run inference, you can try FP8 first; if your plan is to continue pretraining or do full-parameter fine-tuning from this checkpoint, BF16 is the safer starting point. The premise that would change this choice is whether your hardware supports FP8 compute and whether you have a comparable fine-tuning budget. A measurable criterion: fix decoding parameters and run FP8 and BF16 on your own held-out set, comparing task pass rates or exact match; if FP8's drop exceeds a preset tolerance, fall back to BF16 or restrict FP8 to low-risk requests. On the serving stack: the reported path is vLLM 0.32.0 plus the official plugin, or SGLang with a 4-way tensor-parallel reference configuration. If your cluster already runs vLLM and you can pin the version, the plugin path has lower migration cost; if version upgrades are controlled or you want to avoid a plugin dependency for now, the SGLang reference configuration is simpler. The conditions that would change this judgment are how firmly the version can be pinned and how much operations accepts third-party plugins. A measurable criterion: at your target context length, measure per-GPU concurrency, peak memory and decode tokens per second, then decide whether longer context is worth adding GPUs for. On whether to start from this base model: the report explicitly says it runs through the completions endpoint, and that is the fork in the road. If your task requires out-of-the-box instruction following and tool calling, use an existing aligned model first; this checkpoint only adds value when you genuinely need to compare different post-training approaches from the same checkpoint and have the engineering capability for SFT or RL. The measure: compare your own post-trained model against your current model on the same internal task set by pass rate, and do not migrate if it does not win.
Limits and open questions
There are quite a few unknowns. GPU model, memory and single-machine serving scale are not disclosed in the report; all benchmarks are Baidu self-reported, and competitor numbers under Baidu's pipeline differ slightly from what each vendor publishes—for example, DeepSeek lists MMLU-Pro at 68.3 while Baidu measured 69.2, and Tencent lists Hy3 at 65.76 while Baidu measured 66.9—which means this comparison is only good for relative positioning, not cross-vendor conclusions. The 1M context is the server's accepted maximum, and the report gives no quality data at that length; the mid-pack LongBench-v2 score of 48.7 also suggests its long-context ability is not uniformly leading, so developers need to measure memory and latency at their target length themselves. The license terms as relayed by the report, including the revenue threshold and the EU, UK and South Korea exclusions, need to be verified against the license text. As of the catalog check on November 9, 2026, KimiSeek's online model catalog has no ERNIE 5.5 Flash Base entry; any API availability mentioned in the report refers to Baidu, not this site, and for hosted calls please rely on Baidu's official channels—using it through this site would first require an exactly matching catalog entry. In addition, the FlashMask-4 training kernel speedup figures for SM90 and SM100 are also vendor self-reported, and the same release was already covered by Chinese-language media on November 5; what is added here is the architecture, deployment prerequisites and license scope—the parts that affect integration decisions more.