A company that built its AI reputation on keeping models behind a paywall just spent two years handing them out for free. Baidu's ERNIE (文心) family is now one of the most downloaded open-weight model lines in China. The reversal says more about the pressures on China's oldest AI incumbent than about any change of heart.
The reversal, in one timeline
Baidu's large-model (大模型) story moved fast once the open-weight wave broke:
- March 2025: ERNIE 4.5 and the reasoning model ERNIE X1 launched.
- 30 June 2025: Baidu open-sourced the ERNIE 4.5 family under Apache 2.0 — ten models, the largest a vision-language mixture-of-experts with 424 billion total parameters and 47 billion active.
- September 2025: ERNIE X1.1 upgraded the reasoning line; a thinking variant, ERNIE-4.5-21B-A3B-Thinking, briefly topped Hugging Face's global trending chart.
- January 2026: the formal ERNIE 5.0 shipped as a unified multimodal model; a snapshot, ERNIE-5.0-0110, scored 1,460 on the LMArena text board, inside the global top ten.
- May 2026: ERNIE 5.1 followed as a leaner, text-and-reasoning-focused successor.
The throughline: Baidu now gives away the previous generation and sells the frontier through ERNIE Bot, its Qianfan cloud platform, and enterprise deals.
Why Robin Li changed his mind
For years Baidu's founder Robin Li (李彦宏) argued that open models were a poor business. The shift came under direct competitive threat. When DeepSeek and Alibaba's Qwen released strong open-weight models, Baidu's closed ERNIE lost developer gravity. In a January 2026 interview Li reframed the position precisely: the models are "open weight," he said, because the training data needed to reproduce them is not public. The wording matters — Baidu is willing to give researchers the weights, not the recipe.
The strategic logic is defensive. Baidu's moat was always distribution: search, cloud, and a huge base of enterprise customers. If developers standardize on someone else's open model, that moat leaks. Open-sourcing ERNIE 4.5 is a way to keep Baidu inside the workflow even when it is not the one being paid.
What ERNIE 5.0 and 5.1 actually claim
Baidu's own technical report describes ERNIE 5.0 as a 2.4-trillion-parameter mixture-of-experts model that jointly handles text, images, audio and video from a single autoregressive framework, activating under 3% of parameters per token. The company says its language and multimodal understanding is comparable to Gemini 2.5 Pro and GPT-5-high.
ERNIE 5.1, released in May 2026, is the more interesting engineering story. Baidu describes it as a compression, not a scale-up: it inherits ERNIE 5.0's pre-training foundation while shrinking total parameters to roughly a third and active parameters to about half, using an "elastic pre-training" scheme with variable routing sparsity. The headline efficiency claim — pre-training cost about 6% of comparable models at the same scale — is Baidu's own, and the company has published no absolute parameter count for 5.1, only the ratios.
Where it still lags the open field
One distinction keeps biting Baidu. Rivals — DeepSeek, Qwen, GLM, Kimi, MiniMax, Tencent — ship downloadable weights for their current flagships. Baidu open-sources the previous generation only; ERNIE 5.0 and 5.1 weights have never been released. For any team that needs to inspect, adapt or self-host the model they use, that rules out Baidu's frontier entirely.
Benchmark readings also need care. Baidu's launch claims — X1.1 "on par with GPT-5 and Gemini 2.5 Pro," 5.1 scoring 99.6 on AIME26 "second only to Gemini 3.1 Pro" — are vendor evaluations. The only broad third-party leaderboard carrying the model, arena.ai (the rebranded LMArena), placed ernie-5.1 at rank 27 on the text board in a July 2026 snapshot, behind three other Chinese entries, though it sat at rank 5 on the search board.
The bigger corporate picture
The model push sits inside a wider Baidu restructuring. The company reports AI as a distinct management category — AI Cloud, AI Applications, AI-native marketing — and in late 2025 it filed a Hong Kong IPO application for Kunlunxin, its AI chip arm, a move Reuters reported in December 2025. Tying proprietary models to in-house silicon is the same "full stack" logic Alibaba is pursuing, and it signals that Baidu sees models, clouds and chips as one contest rather than three.
What readers can do now
- If you build on Chinese models, benchmark ERNIE 4.5 (open weights) on your workload before paying for 5.x; the gap may be smaller than the marketing implies.
- If you track the industry, watch whether Baidu ever releases 5.x weights. A continued "previous-gen only" policy would confirm models are now a cloud-and-search retention tool, not a product.
- If you use Baidu search or cloud, expect ERNIE to be baked deeper into results and enterprise tools through 2026; test the new agentic and search features directly.
Honest limitations
Core facts (ERNIE 4.5 launch March 2025; open-sourced 30 June 2025 under Apache 2.0, ten models, up to 424B total / 47B active; ERNIE-4.5-21B-A3B-Thinking top Hugging Face trending Sept 2025; ERNIE X1.1 9 Sept 2025; ERNIE 5.0 preview Baidu World 13 Nov 2025 and formal release 22 Jan 2026; ERNIE-5.0-0110 LMArena 1,460 Jan 2026; ERNIE 5.1 May 2026; Kunlunxin HK IPO filing reported by Reuters Dec 2025) come from Baidu's releases, Baidu Baike, AI Wiki, AICHatDaily and Reuters. Parameter counts, the "<3% active" claim and the "6% pre-training cost" figure are Baidu's own technical-report claims, not independently audited. The arena.ai July 2026 ranking is a third-party snapshot. Robin Li's "open weight" phrasing is from a January 2026 interview. This article is current to 29 September 2026 and is not investment advice.