ERNIE 5.5 Flash Preview
Baidu·
Overview
First post-trained model built on the open ERNIE 5.5 Flash Base, described by Baidu as an intermediate checkpoint from post-training: 280B total / 14B active MoE, text-only, 1M-token context with deep-thinking mode. API-only for now on Baidu Qianfan and OpenRouter (listed there 2026-11-10; Baidu's post is dated 2026-11-11). Baidu says the weights will be open-sourced under Apache 2.0, with no date; until they are published this entry stays cloud-only. Qianfan lists prices in RMB only and the OpenRouter price could not be confirmed on a primary page, so no pricing block is carried (checked 2026-11-13). All benchmark figures are Baidu's own, at the highest available reasoning setting.
Capabilities and innovations
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| Humanity’s Last Exam (no tools) | 46.4% | vendor 'HLE (w/o tools)' 46.4 on the text-only subset, highest reasoning setting, recorded for comparability with the no-tools HLE used here. Baidu's footnote names Gemini 3.1 Flash-Lite as the judge, which is not the benchmark's standard grader. Read from the benchmark table image linked in the launch post. |
| Terminal-Bench 4.0 | 46.5% | vendor Terminal-Bench 4.0 46.5 from the Claude Code 2.1.246 harness, highest reasoning setting, launch post table. Vendor figure from a non-reference harness; no Vals AI run exists yet (checked 2026-11-13). 4.0 is its own generation and not comparable with the 2.x or 3.0 fields. |
| DeepSWE v1.1 | 71.2% | vendor DeepSWE v1.1 71.2, explicitly labelled v1.1, mini-swe-agent 2.4.6 'with anti-hacking policies', launch post table. No earlier Baidu model appears in the table, so there is no vendor baseline row to check. Competitor cells marked with an asterisk are Baidu's own re-runs and are not entered anywhere from this source. Datacurve's own board was not checked. |
- GPQA Diamond: not reported by the vendor
- SWE-bench Verified: not reported by the vendor
- SWE-bench Pro: not reported by the vendor — Baidu reports DeepSWE v1.1, ProgramBench and NL2Repo-Bench instead.
- MMLU: not reported by the vendor
- MMLU-Pro: not reported by the vendor
- Terminal-Bench 2.x: not reported by the vendor — Only Terminal-Bench 4.0 is reported, recorded in terminal_bench_4.
- Terminal-Bench 3.0: not reported by the vendor
- MMMU-Pro: not reported by the vendor — Not applicable — the model is text-only.
Architecture and hardware
- Parameters
- 280B total, 14B active per token (MoE)
Links
More from Baidu
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.