Bottom Line First
Baichuan AI entered a relatively quiet period after releasing Baichuan 4 in late 2025. Compared to Qwen 3.6, DeepSeek V4, and Kimi K2.6 from the same timeframe, Baichuan’s presence on benchmark leaderboards has significantly diminished. But this doesn’t mean Baichuan is “falling behind” — it is pursuing a differentiated vertical industry route, just shifting its volume strategy from “leaderboard racing” to “deep cultivation.”
Technical Roadmap: The Real Level of Baichuan 4
| Dimension | Baichuan 4 | Qwen 3.6 35B | DeepSeek V4 | Kimi K2.6 |
|---|---|---|---|---|
| Parameter Scale | ~13B (MoE) | 35B A3B MoE | 236B/27B MoE | Undisclosed |
| Chinese Reasoning | First Tier | First Tier | First Tier | First Tier |
| Coding Ability | Mid-Upper | Top | Top | Mid-Upper |
| Open Source Level | Partial | Full | Full | Not Open |
| Pricing Strategy | Low | Extremely Low/Free | Extremely Low | Medium |
Baichuan 4 uses a MoE (Mixture of Experts) architecture, achieving performance close to 70B dense models at a 13B parameter scale. This technical choice has its rationale — pursuing cost-effectiveness under compute constraints.
But the problem is also obvious: when competitors push parameter scales to 35B or even 236B, the ceiling of 13B MoE is simply not enough.
Open Source Strategy: From “Full Open” to “Selective”
Baichuan was among the first Chinese model providers to embrace open source. Early models like Baichuan-7B and Baichuan-13B gained significant attention on Hugging Face. But starting with Baichuan 4, the open source strategy noticeably contracted:
- Baichuan 4 only released a reasoning API, not model weights
- The community cannot fine-tune or self-deploy
- Lost organic traffic and secondary propagation on Hugging Face
Compared to the Qwen team open-sourcing weights for nearly every model, and DeepSeek being even more aggressive with open source, Baichuan’s choice appears conservative. This directly caused a break in the community ecosystem — developers no longer build applications on top of Baichuan models.
Vertical Industries: Baichuan’s “Hidden Thread”
Baichuan’s real growth point may not be general benchmarks, but vertical industries. Cross-verified from multiple sources:
- Healthcare: Baichuan has partnered with multiple top-tier hospitals to develop medical Q&A and diagnostic assistance systems
- Education: Fine-tuned versions of Baichuan 4 for education scenarios have been piloted in several key schools
- Finance: Knowledge Q&A systems co-developed with brokerages have been deployed
The advantage of this route is large profit margins and strong customer stickiness; the disadvantage is lack of public visibility, making it harder to build a brand flywheel.
Landscape Judgment
| Provider | Current Strategy | Risk | Opportunity |
|---|---|---|---|
| Qwen | Full open source + ecosystem binding | Rising open source compliance costs | Becoming the global Chinese foundation model |
| DeepSeek | Extreme cost-effectiveness + Agent integration | Business model sustainability | Becoming an AI infrastructure layer player |
| Kimi | Closed source + long-context differentiation | Single technical route dependency | Entering office scenarios |
| Baichuan | Vertical industry deep cultivation | General capability gap widens | Industry Know-how barrier |
Baichuan is not “falling behind,” but has actively chosen a different path. But this path has one prerequisite: can vertical industry revenue sustain continuous model R&D investment? If Qwen and other providers also start doing industry customization, Baichuan’s differentiation advantage will be quickly diluted.
Action Recommendations
- If you are selecting a general foundation model: Prioritize Qwen 3.6 and DeepSeek V4 at this stage — open source ecosystem and cost-effectiveness are both more mature
- If you are doing industry implementation: Reach out to Baichuan about industry solutions — their experience in healthcare and education is worth referencing
- If you are making investment judgments: Watch Baichuan’s next funding round and industry customer renewal rates — these are key indicators of whether their strategy is working
Baichuan’s story is not over. In the second half of AI models shifting from “parameter racing” to “scenario racing,” whether the vertical route can succeed, Baichuan will be one of the most worth-observing cases.