ERNIE-5.5-Flash vs Qwen 3.8 Flash: A Base Checkpoint Meets a Finished Model

Put ERNIE-5.5-Flash next to Qwen 3.8 Flash and the first honest observation is that they are not the same kind of object. What Baidu has actually published is ERNIE-5.5-Flash-Base, the checkpoint at the end of pre-training — before mid-training, before instruction tuning, before any preference work — so it continues text rather than answering requests, and it is served through a completions endpoint, not a chat one. Qwen 3.8 Flash, which Alibaba shipped on 26 August 2026, is a finished multimodal model that takes text, images and video and returns text, at $0.15 per million input tokens and $0.47 per million output. One is a starting point you still have to train; the other is a product you can call this afternoon. Nearly every number below changes meaning depending on which of those two things you think you are comparing, which is why it is worth settling before the spec sheet rather than after.

What Baidu shipped, and what it has not said

The release is quiet in the specific way that Chinese open-weight drops often are: the artifacts arrive before the story does. The baidu/ERNIE-5.5-Flash-Base repository was created on Hugging Face on 21 October, sat near-empty for a week, then took its 128 FP8 weight shards between 2 and 3 November. The architecture and benchmark diagrams landed on 4 November, and the card itself was still being revised this morning. Baidu's own release page is dated today, 5 November, and the only discussion it has generated outside that page is a forum thread and a WeChat post. There is no press release, no launch event, and no press coverage to speak of.

That matters for what you can and cannot trust. Knowable from the repository: the architecture, the parameter counts, a complete vendor benchmark table with an explicitly named harness, working SGLang and vLLM serving recipes, the licence text, and the fact that a BF16 sibling exists alongside the FP8 one. Not knowable: whether Baidu intends to serve it as an API — Qianfan's price sheet, last updated 18 October, lists ERNIE 5.1, 5.0 and 4.5 Turbo, and no ERNIE 5.5 entry of any kind — what an instruction-tuned version of it scores, and whether any independent evaluator will reproduce a single number in Baidu's table.

The magazine spec, for the record: 280B total parameters with 14B active per token, pre-trained from scratch on 33.5T tokens, 43 layers of which one is a dense FFN and 42 are MoE, 512 routed experts with top-10 routing plus one shared expert, and a 201,216-token vocabulary. The released config enables a 1,048,576-token context, though training only ever saw 131,072. There is also a 0.36B multi-token-prediction head that is not counted in either parameter figure.

The spec contrast, on the dimensions that actually decide

• What you receive — ERNIE-5.5-Flash-Base is a checkpoint and nothing else: weights you download, a /v1/completions endpoint, and no instruction-following behaviour. Qwen 3.8 Flash is a served model with tool calling, reasoning mode, structured outputs and a 131,072-token maximum response.

• Size — 280B total / 14B active for ERNIE against 125B total / roughly 6B active for Qwen 3.8 Flash, which also carries a 51B n-gram embedding table and a ~4B MTP head, putting about 180B of parameters on disk.

• Attention design — ERNIE uses 43 layers split 37 heavily-compressed-attention (HCA) layers plus 6 MLA layers with sparse attention; Qwen 3.8 Flash uses a Gated DeltaNet plus Qwen Sparse Attention hybrid across 48 layers. Both are bets that classic full attention is the wrong thing to pay for at Flash-tier prices, made from opposite directions.

• Context — 1M tokens claimed on both sides, from different places: ERNIE trained at 128K and enables 1M only at fine-tune and inference time, while Qwen 3.8 Flash is natively 262K and reaches 1M through YaRN extension.

• Modality — ERNIE-5.5-Flash-Base is text only, in English and Chinese. Qwen 3.8 Flash accepts text, images and video, and its vendor table includes AndroidWorld, OSWorld 2.0, MathVision and LVBench results that a text-only checkpoint cannot produce.

• Weights on the Qwen side — the production Qwen 3.8 Flash is API-only, but its open-weight sibling Qwen3.8-Flash-Next is published under the Qwen Community License 1.0, so the two sides are more comparable than the API-versus-weights framing first suggests.

• Serving — Baidu ships a working recipe: the lmsysorg/sglang:dev-ernie-55-flash image with --tp 4, or vLLM 0.32.0 plus the baidu/vLLM-ERNIE plugin with data-parallel size 4. Qwen 3.8 Flash needs no serving recipe because Alibaba serves it for you.

• Price — no published per-token price for ERNIE-5.5-Flash-Base anywhere, because there is no hosted endpoint to price; $0.15 input / $0.47 output per million tokens for Qwen 3.8 Flash, with cached input reads at $0.0184.

• Licence — the ERNIE 5.5 Flash Base Model License Agreement v1.0, a custom agreement worth reading in full before you build a product on it, against Qwen Community License 1.0 on the open Qwen side.

The benchmark rows that look comparable and are not

Here is the trap this comparison sets. Both vendors publish a LiveCodeBench number, and the two numbers are 40.1 and 91.9.

They are not measuring the same thing, and the gap between them is not a capability gap. Baidu's 40.1 is Pass@1, one-shot, on the raw pre-trained checkpoint — the model is being asked to complete code continuations having never been taught to follow an instruction. Alibaba's 91.9 is LiveCodeBench v6, a different and later benchmark version, reported for a model that has been through a full post-training pipeline. The GPQA pair has the same problem: 60.2 for ERNIE at five-shot exact match on a base checkpoint, 91.7 for Qwen on GPQA Diamond after instruction tuning. Even the MMLU comparison, 93.3 against a model that publishes no MMLU, has nothing to sit beside.

This is not a knock on either model. It is a statement about what a base-model evaluation is for. A 40.1 on LiveCodeBench from a checkpoint that has never seen an instruction is a statement about how much coding structure pre-training absorbed, and it is a reasonable number on that reading. It tells you nothing about how the same weights would score after someone fine-tuned them. Nobody, including Baidu, has published that number.

The one table that can be read at face value is Baidu's own, because all four columns were run through the same pipeline, prompts and few-shot settings. Against DeepSeek-V4-Flash-Base (284B-A13B), Hy3-preview-Base (295B-A21B) and MiMo-V2.5-Base (310B-A15B), ERNIE-5.5-Flash-Base takes best-in-row on 24 of 30 benchmarks, including MMLU-Pro at 78.4 against 69.2, HumanEval+ at 82.3 against 69.5, and RepoQA at 91.7 against 81.0. It loses ChineseSimpleQA to DeepSeek, FullStackBench and BBH chain-of-thought to Hy3, and both HELMET-Rerank and InfiniteBench to DeepSeek. Those are vendor-run numbers on a vendor-chosen harness with no independent replication — but they are internally comparable, which is more than can be said for any cross-vendor pairing in this article.

What each route costs, with the arithmetic shown

Qwen 3.8 Flash is the easy one to price. Take a support-and-extraction workload of 1,000 requests a day, averaging 6,000 input tokens and 800 output tokens: 6M input and 0.8M output tokens daily, or 180M and 24M across a 30-day month. At $0.15 and $0.47 per million that is $27.00 plus $11.28 — $38.28 a month, with cached reads at $0.0184 per million shaving the recurring system-prompt portion further.

ERNIE-5.5-Flash-Base has no equivalent line item, so the only honest cost model is hardware. The FP8 repository is 128 shards of roughly 2.25 GB each, about 288 GB of weights; the BF16 sibling doubles that. Baidu's own SGLang launch command shards it four ways with --tp 4, which puts about 72 GB of weights on each GPU before any KV cache exists — workable on an 80 GB card only if you never use the context window, and comfortable on H200-class cards. Rent four of those at a mid-range $2–3 per GPU-hour and the node costs roughly $5,800–8,600 a month running continuously, before you have paid anyone to fine-tune it, evaluate it, or keep the serving stack current.

That is on the order of 150 to 225 times the Qwen API bill in the example above, and the comparison is deliberately unfair in one direction: the two are not substitutes until someone post-trains the ERNIE checkpoint. The point of the arithmetic is not that self-hosting is expensive — it is that the fixed cost is real and it arrives before the model can answer a single instruction, so the volume at which owning it beats renting something finished is much higher than a per-token comparison suggests.

Long context is where the two designs converge

Both models claim a million tokens, and both get there by refusing to pay full attention over a million positions. ERNIE's HCA layers cache a single entry per 128 tokens, so 37 of its 43 layers grow their KV cache 128 times more slowly than a conventional layer would; the six MLA layers keep full-resolution access for the tokens that need it. Qwen's GDN-plus-QSA hybrid makes the same kind of trade from the sparse-attention side, and its 6B active parameters keep per-token compute low regardless of what the attention is doing.

For a reader, the practical difference is not which design is cleverer — both are credible — but that only one of the two comes with a price for actually using that context. A million-token prompt on Qwen 3.8 Flash has a number attached to it today. On ERNIE-5.5-Flash-Base the same prompt has a GPU-hour estimate and an unanswerable question about how a base model behaves at that length without instruction tuning.

Who should pick which

Pick Qwen 3.8 Flash if you need a model that follows instructions now, if your inputs include screenshots, documents or video, if you want the long-context bill to be somebody else's capacity-planning problem, and if $0.15/$0.47 per million tokens fits the unit economics of what you are building. It is two months old, it has had third-party trackers pick it up since launch, and the vendor numbers — SWE-bench Pro 62.5, GPQA Diamond 91.7, CoWorkBench 73.9 — are claims rather than verified results, so treat them as an upper bound until your own eval says otherwise.

Pick ERNIE-5.5-Flash-Base if you have a post-training capability and a reason to own the model rather than rent one: a domain whose data you cannot send to an API, a Chinese-and-English workload where CMMLU 90.0 and ChineseSimpleQA 73.0 are the numbers you care about, an RL research programme that wants a 280B-A14B starting point, or a latency budget that justifies a dedicated four-GPU node. You also need to be comfortable that the licence is a custom Baidu agreement, not Apache, and that the model does nothing useful until your team makes it useful.

Pick neither, yet, if what you actually want is a proven general-purpose model. Qwen 3.8 Flash is a finished product with no independent benchmark score, and ERNIE-5.5-Flash-Base is an unfinished model with a vendor benchmark table. Neither is the safe default that a two-year-old flagship would be.

Running the experiment without betting a roadmap on it

The awkward part of this matchup is that the two sides are reached in completely different ways, and a routing layer helps with exactly one of them. Qwen 3.8 Flash is available through OrcaRouter as qwen/qwen3.8-flash, at provider list price with 0% markup — the same $0.15 and $0.47 the vendor charges, passed through, with nothing added for the hop. ERNIE-5.5-Flash-Base is not something we serve: it is a weights release, and no hosted endpoint for it exists to route to, here or anywhere else we can find.

What a single endpoint is genuinely good for in this situation is the part that comes after the experiment. The realistic path for an ERNIE 5.5 Flash is that a team post-trains the base checkpoint, and the question then becomes whether the tuned model beats Qwen 3.8 Flash on your own traffic. With both behind one key, that is a model-id change rather than a second contract and a second SDK, the routing DSL can split production traffic between them while you decide, and automatic failover means an unproven winner cannot take the whole feature down on its first bad day. Model fusion gives you the third option — a panel answering together — which is often the cheapest way to find out whether a new tune is actually better than the incumbent or merely different.

What would change this answer

Three things, in rough order of likelihood. An instruction-tuned ERNIE 5.5 Flash would turn this from a category mismatch into a real head-to-head, and the base numbers suggest it would be competitive — 82.3 on HumanEval+ and 91.7 on RepoQA from an untuned checkpoint is a strong starting position. An ERNIE 5.5 entry on Qianfan's price sheet would give the self-hosting arithmetic something to compare against and tell you whether Baidu intends to sell this model or give it away. And a third-party score for either model would let a reader stop taking vendor tables on trust, which is the state this pairing is stuck in today: Qwen 3.8 Flash has no independent benchmark result to its name, and ERNIE-5.5-Flash-Base has no row on any neutral leaderboard at all.

Until two of those land, "which is better" is not a question anyone has measured. What can be answered today is narrower and more useful: Qwen 3.8 Flash is a model you can call, and ERNIE-5.5-Flash-Base is a model you can build on. Those are different jobs, and no benchmark table is going to settle which one you have.