ERNIE-5.5-Flash vs GLM-5.3 Flash: Two Open Repos, Two Very Different Deals

Both ERNIE-5.5-Flash and GLM-5.3 Flash are downloadable this week, both are sparse mixture-of-experts models in the same weight class, and both fit on roughly the same shelf of GPUs in FP8. That is where the similarity stops. GLM-5.3 Flash, released by Z.ai on 26 August 2026, is a finished multimodal model under the MIT licence with a published price and a served endpoint. ERNIE-5.5-Flash-Base, released by Baidu on 5 November 2026, is the end-of-pretraining checkpoint of a model that has not been post-trained, under a bespoke licence, with no price and no endpoint. Comparing them as rival open models is reasonable. Comparing them as substitutable ones is not, and the gap is legal and procedural before it is technical. Both GLM variants are billed at provider list prices so the free-tier variant costs nothing to try.

What is actually in each repository

GLM-5.3 Flash is a 320B-parameter model with 18B active per token, 45 layers, and a hybrid of linear and sparse attention that Z.ai says cuts attention compute roughly threefold and shrinks the KV cache about 4.4 times relative to the text-only GLM-5.3. The context window is 1,048,576 tokens, output is capped somewhere between 128K and 131K depending on the host, and the model takes text, images and video in with text out. The FP8 weights on Hugging Face run to roughly 328GB.

ERNIE-5.5-Flash-Base is 280B total with 14B active, 33.5T pretraining tokens, 43 layers — one dense plus 42 MoE — with 37 layers on compressed HCA attention and six on MLA plus DeepSeek-style sparse attention. Hidden size 4,096, 64 heads, 512 routed experts at top-10 plus one shared expert, vocabulary 201,216. Training context 128K, with 1,048,576 enabled for fine-tuning and inference in the released config. FP8 and BF16 variants, both on Hugging Face.

• Total parameters — ERNIE 280B vs GLM 320B

• Active per token — 14B vs 18B

• Layers — 43 vs 45

• Context — 1M for inference vs 1M

• Modality — text only vs text, image and video in

• Licence — ERNIE 5.5 Flash Base License v1.0 vs MIT

• Post-training — none vs instruction-tuned

• Price — none published vs $0.15 input / $0.50 output

The parameter gap is smaller than the naming suggests and runs the opposite way from what you might assume: ERNIE activates fewer parameters per token and has fewer total, which on paper makes it the cheaper of the two to serve. GLM's 328GB FP8 footprint is the concrete number; ERNIE at 280B lands in the same order of magnitude, before KV cache.

The licence is the first fork in the road

MIT is not a marketing detail. It means commercial use, modification, redistribution and derivative models with no field-of-use terms to negotiate and no separate agreement to sign. A team can fine-tune GLM-5.3 Flash on proprietary data and ship the result without a conversation with Z.ai's legal department.

ERNIE-5.5-Flash-Base ships under the ERNIE 5.5 Flash Base Model License Agreement v1.0. That is a bespoke agreement, and the contrast worth knowing is historical: Baidu released the ERNIE 4.5 family under Apache 2.0 in 2025. Moving to a custom licence for the 5.5 base is a deliberate tightening, not an oversight, and it lands at exactly the moment the weights become genuinely useful — a base checkpoint is far more likely to be fine-tuned and redistributed than an instruction-tuned model is.

For academic and research use the practical difference may be nil. For anyone building a product, it is the first document to read and possibly the reason to choose the other repository. Weights you can download are not the same as weights you can build on.

Modality decides more workloads than benchmarks do

GLM-5.3 Flash accepts video. That single capability removes it from comparison with ERNIE for a large class of 2026 workloads — screen recordings, product footage, surveillance clips, lecture capture. GLM also handles images, which covers invoices, charts, screenshots and the entire document-processing category.

ERNIE-5.5-Flash-Base is a language model. Text in, text out. If the post-trained ERNIE 5.5 Flash ever ships with multimodal input, Baidu has not said so; the release describes a text model and the benchmark suite covers knowledge, reasoning, math, code, multilingual and long context. Nothing in the published material suggests vision.

That asymmetry is worth stating plainly because it is easy to lose inside a spec comparison. A 280B text model and a 320B multimodal model are not alternatives in a procurement decision; they are answers to different questions. The only overlap is text-only generation, and in that overlap the finished model wins by default, because it exists.

Where GLM-5.3 Flash is weaker than its launch suggested

Honesty runs both ways. GLM-5.3 Flash's published strengths are real — DeepSWE 63.4 against GLM-5.2's 46.2, Terminal-Bench 2.1 at 84.3, MMLU-Pro 86.1, MMMU-Pro 86.0, HLE with tools 55.3 — and Artificial Analysis puts its coding index at 71.5 and its general intelligence index at 41.8. But the same evaluations show a model with narrow, deep weaknesses: ProgramBench scores around zero to 1.5, SRE Bench 0.8, Legal Agent 6.7. Its first-token latency is also notably slow for a Flash-tier model — 6.36 seconds at p50 across OrcaRouter's seven-day traffic — against roughly 1.7 seconds for the larger GLM-5.

None of that makes ERNIE the better choice. It makes GLM a known quantity — a model whose failure modes have been measured by third parties. ERNIE's failure modes have not been measured at all, because there is no instruct model to measure. In an evaluation-driven decision, unknown failure modes are worse than documented ones, even when the documented ones are bad.

Fine-tuning paths: both exist, one has a head start

Baidu did the right thing on tooling. The ERNIE model card includes SGLang and vLLM examples, and NVIDIA's Megatron-Bridge repository added an ERNIE 5.5 example that covers fine-tuning. For a base checkpoint, that is the difference between a research artifact and a usable starting point.

GLM-5.3 Flash has been downloadable since late August and has accumulated the ecosystem mileage that comes with three months of production use: quantizations, provider integrations and bug reports. When a kernel breaks at 2am, the GLM issue has already been filed by someone else. That is not a benchmark, but it is a real cost of adoption and it should be in the decision.

For anyone who wants to try GLM-5.3 Flash without committing a GPU cluster, OrcaRouter routes z-ai/glm-5.3-flash at the provider's list price with 0% markup, and there is a free-tier variant in the catalogue as well. The same catalogue has no ERNIE entry among its 216 model IDs — no hosted path, no price, nothing to route. That is not a criticism of Baidu's release; it is the practical consequence of shipping weights before a product.

Which repository to clone

Clone GLM-5.3 Flash if you need a multimodal open model you can put behind an API, fine-tune commercially, and compare against published third-party evaluations. MIT plus a served endpoint plus three months of production mileage is a strong combination, and the documented weak spots are narrow enough to route around.

Clone ERNIE-5.5-Flash-Base if you are doing continued pretraining, domain adaptation or post-training research, particularly in Chinese, and you are prepared to read a bespoke licence before you invest in the pipeline. The 280B-A14B shape is efficient, the 33.5T-token budget is serious, the architecture is documented in more detail than most base releases, and the Megatron-Bridge example means the first fine-tune is not a research project of its own.

Do not clone either expecting a drop-in replacement for the other. One is a finished product with a licence that permits almost anything; the other is a strong starting point with a licence that permits less and a product that does not exist yet. The interesting question is what Baidu ships in the post-trained model and at what price — because if it lands near $0.15/$0.50 with 14B active parameters, this stops being a comparison of two repositories and becomes a genuine contest between two open Flash models. Until then, the repositories are not on the same shelf at all.