
Hy4 flash-Base vs ERNIE 5.5 Flash Base: The Only Fair Comparison Either Vendor Published
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 138 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 175 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 89 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 97 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-079 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 88 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 86 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 94 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 120 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 68 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 92 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 44 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 195 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 116 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
Two Chinese labs published pre-trained checkpoints twelve days apart, and for once the comparison table writes itself: Hy4 flash-Base against ERNIE 5.5 Flash Base, the same kind of artefact from the same stage of the pipeline, evaluated the same way. Tencent published Hy4 flash-Base on 17 November 2026 alongside the instruction model. Baidu published ERNIE 5.5 Flash Base on 5 November 2026, and at the time of writing it is still the only ERNIE 5.5 artefact that exists — the post-trained chat model has not shipped.
That difference in what surrounds the checkpoint is most of what makes this matchup worth a page. Comparing the two bases tells you how the pre-training went. Comparing the two releases tells you something less flattering about both: one company shipped the raw material and the finished product on the same day, the other shipped the raw material and a promise.
Why base checkpoints are the honest comparison
Almost every model comparison you read is quietly unfair, because it sets one vendor's finished, instruction-tuned model against another vendor's finished, instruction-tuned model, each evaluated on its own harness with its own baselines. Base checkpoints remove one of those variables. Neither model has been post-trained, neither has a chat template that shapes the output format, and both vendors publish few-shot tables on standard academic benchmarks rather than custom agentic suites.
What remains — the harness, the shot counts, the subsets — is still vendor-specific, so nothing here is a controlled experiment. But a base-versus-base table is a comparison of pre-training runs, and pre-training is the expensive part. It is the closest thing to an apples-to-apples number these two companies have published about each other.
The two checkpoints, side by side
• Parameters — Hy4 flash-Base 220.3B total, 12B active per token; ERNIE 5.5 Flash Base 280B total, 14B active per token.
• Where the parameters sit — Hy4 flash-Base carries 155.5B in the 42-layer backbone, 63.5B in Engram n-gram embedding tables, 0.43B in a vision encoder and 0.85B in a DFly2 drafter; ERNIE's 280B is a more conventional MoE spread, plus a 0.36B ECHO module excluded from the headline count.
• Depth and routing — 42 layers (one dense, 41 MoE) with 512 routed experts plus one shared, 16 active, hidden size 3,072, for Hy4 flash-Base; 43 layers (one dense, 42 Multi-Router Latent MoE) with 512 routed experts plus one shared, top-10, hidden size 4,096, for ERNIE.
• Context — 1,048,576 tokens configured on both. Tencent trained Hy4 flash directly at that length; Baidu trained ERNIE at 128,000 and set the released config to 1,048,576 for fine-tuning and inference, which is a capability of the configuration rather than a measured behaviour.
• Modality — text and image in, text out, on both. Hy4 flash-Base ships a SigLIP2-derived vision encoder; ERNIE 5.5 Flash Base's card describes text input only on the model-link table.
• Licence — Tencent Hy License v1.0, an Apache 2.0 derivative with a MaaS clause at $50 million revenue or 100 million MAU; ERNIE 5.5 Flash Base Model License v1.0, which excludes the European Union, the United Kingdom and South Korea as territories entirely.
The licence row is not a footnote. A European team cannot use the ERNIE checkpoint at all under its grant, and no amount of benchmark advantage changes that. A team of any size can use Hy4 flash-Base, and a team above $50 million in MaaS revenue needs to send an email before it does.
What Baidu's table says, and what it costs to believe it
Baidu's comparison set is the right one for a base release: ERNIE 5.5 Flash Base against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, on one pipeline the card describes as shared across all four. Every figure below is Baidu-reported and unreproduced.
• MMLU — 93.3 ERNIE, 89.4 DeepSeek-V4-Flash-Base, 87.5 Hy3-preview-Base, 87.6 MiMo-V2.5-Base.
• MMLU-Pro — 78.4 against 69.2, 66.9 and 65.7. The largest gap in the set.
• GPQA — 60.2 against 55.9, 50.2 and 51.4.
• HumanEval+ — 82.3 against 69.5, 78.0 and 74.4.
• LiveCodeBench — 40.1 against 31.9, 33.5 and 32.4.
• LongBench-v2 — 48.7 against 44.5, 39.8 and 44.6.
• BBH with chain-of-thought — 89.2 against a Hy3-preview-Base 90.0 and MiMo-V2.5-Base 89.8. A three-way tie inside noise, whatever the bold formatting implies.
• ChineseSimpleQA — 73.0 against DeepSeek's 75.3. A loss, on the benchmark closest to Baidu's home market.
The table loses rows, which is why it reads as a table somebody built rather than one somebody picked. Note what is absent from it: Hy4 flash-Base. Tencent's checkpoint did not exist on 5 November, so no direct row between the two models in this comparison has ever been published by anybody.
What Tencent's table says, including the disclaimer
Tencent's base table is five-way — Hy4 flash-Base, Hy4 Base, Hy3 preview Base, DeepSeek-V4-flash-Base and DeepSeek-V4-Pro-Base — and it also loses rows. Vendor-reported, unreproduced:
• MMLU-Pro — 71.1 for Hy4 flash-Base against 73.9 for Hy4 Base, 65.8 for Hy3 preview Base, 68.6 for DeepSeek-V4-flash-Base and 73.4 for DeepSeek-V4-Pro-Base.
• SimpleQA — 37.5 against 57.2, 26.4, 31.3 and 52.1. The largest loss on the card.
• C-Eval — 92.0 against 95.0, 90.1, 91.9 and 93.4.
• CMMLU — 91.6 against 93.1, 89.6, 89.8 and 90.7.
• AGIEval Chinese — 90.6 against 91.2, 87.8, 88.5 and 89.8.
• Belebele — 93.8 against 95.0, 94.2, 94.4 and 95.2.
• INCLUDE — 82.8 against 88.1, 84.3, 84.0 and 84.3.
• MMMU-Pro — 52.0 against 59.6 for Hy4 Base. The multimodal rows exist only for the two Hy models.
Read as a whole: Hy4 flash-Base is a smaller, cheaper model than its siblings and it loses to them on most rows, which is exactly what a 12B-active model should do against a 49B-active one. The interesting rows are the multilingual ones, where it holds within a point or two of much larger models, and SimpleQA, where it falls twenty points behind Hy4 Base — a fact-retrieval gap that says something about how much of that capability lives in parameter count rather than training data.
Tencent also volunteers a caveat Baidu's card does not: because the base model was exposed to a large amount of synthetic instruction data during mid-training, it "occasionally fails to follow the expected answer format in few-shot evaluations, which may distort the scores." A vendor telling you its own eval harness may be measuring format compliance is rare enough to be worth the sentence.

The two tables cannot be laid on top of each other
Here is the trap in this comparison, and it is worth stating plainly because most of the internet will not. Both vendors publish MMLU-Pro. Baidu reports 78.4 for ERNIE 5.5 Flash Base. Tencent reports 71.1 for Hy4 flash-Base. It is tempting to subtract and report a 7.3-point win for Baidu. That subtraction is not defensible.
Different shot counts, different prompt formats, and in Hy4 flash-Base's case an acknowledged format-following problem that can depress a macro-averaged score. The one number in both tables that comes with a shared definition — Belebele on a five-language subset — reads 93.8 for Hy4 flash-Base and is simply absent from Baidu's table. There is no overlap rigorous enough to subtract.
What you can say is narrower and still useful: each vendor's checkpoint beats the third-party DeepSeek baselines each of them chose, on most rows, by margins their own harnesses produced. Two independent-ish claims in the same direction is stronger evidence than either alone and weaker than a shared benchmark. That is the honest ceiling on what these two cards support.

One release is finished; the other is a promise
This is the asymmetry that matters more than any row above, and it is not about benchmark quality.
Tencent published three artefacts on 17 November: the instruct model, its FP8 build, and the base checkpoint. A team that wants a chat model today can download one, load it in vLLM with the pinned vllm/vllm-openai:hy4-flash image, and serve it. The base checkpoint is there for people who want to fine-tune, and the card points at a fine-tuning guide.
Baidu published one artefact on 5 November. ERNIE 5.5 Flash Base is pre-trained weights. There is no instruction-tuned ERNIE 5.5 Flash to download, no API endpoint documented under that name in the release material, and the card describes the post-trained model only as something that "is supposed to follow." A team that downloads ERNIE 5.5 Flash Base today is committing to doing the post-training themselves, on 280B of weights, before they have a model that can answer a question in a usable format.
For a research group that is precisely the interesting artefact. For a product team it is not a model, it is a starting point with an unbounded amount of work attached, and the correct comparison is not against Hy4 flash-Base but against Hy4 flash — which costs money per token and works now.

Where a router fits into a base-checkpoint decision
Base checkpoints are not what a router serves, and it would be dishonest to suggest otherwise. OrcaRouter carries neither Hy4 flash-Base nor ERNIE 5.5 Flash Base, and neither vendor's instruct model is on the catalogue today. What the catalogue does carry is tencent/hy4-preview — the 770B/49B model that Hy4 flash was distilled from — at $0.834 per million input tokens and $2.501 per million output, with a $0.042 cache read.
That matters for a specific decision. If you are weighing whether to fine-tune a 220B base checkpoint, the first question is whether the family's behaviour suits your task at all, and the preview model answers it in an afternoon for the price of the tokens you spend testing. Doing that before you provision eight GPUs for a tensor-parallel FP8 deployment is the cheapest step in the whole project. The provider list price passes through at 0% markup, so a Tencent price change reaches the catalogue the day it happens rather than at the next rebuild, and automatic failover means an evaluation job does not die because one upstream had a bad hour.
If the fine-tune works, you own the weights and the router stops being part of that path. If it does not, you have spent an afternoon rather than a quarter finding out.
How to read this matchup
If you want a model that can do work, this page is the wrong page: Hy4 flash is the finished model and ERNIE 5.5 Flash does not have one yet. If you want a pre-trained base to build on, the two candidates are closer than the marketing suggests and further apart than a benchmark subtraction would indicate.
ERNIE 5.5 Flash Base is larger, trained longer — 33.5 trillion tokens on the Muon optimizer — and reports stronger scores on its own harness, including an MMLU-Pro figure that no Tencent model on the comparison card matches. Hy4 flash-Base is smaller, cheaper to serve at 12B active, natively multimodal, ships with a vision encoder and a speculative-decoding drafter, and comes with a licence that does not exclude three jurisdictions.
The deciding factor is likely to be neither capability nor licence but the twelve days between them. Baidu's checkpoint has had almost two weeks for the fine-tuning community to produce something; Hy4 flash-Base has had one. Watch whether an instruction-tuned ERNIE 5.5 Flash appears, and watch whether anyone publishes a LoRA for Hy4 flash-Base. The first would make this a real comparison with a shared benchmark most likely to follow; the second would tell you which base the practitioners actually preferred.
Compared in this article1
Detected from this article
