Small models occupy the market: Liquid AI 8B MoE + Hy3 reaches the top record

Core conclusion

In the last week of May 2026, two seemingly unrelated AI events pointed to the same trend: Small and cheap models are becoming the mainstream choice in the AI industry.

On the one hand, Liquid AI released LFM2.5-8B-A1B - a MoE edge model with only 8B total parameters and 1B activation parameters, achieving tool calling capabilities comparable to large models on consumer-grade hardware. On the other hand, Tencent's open source model Hy3 mysteriously topped the OpenRouter usage rankings. Users are crazy about using it not because of the best quality, but because the price of $0.066/million tokens is low enough.

Two signals are superimposed: the AI ​​industry is shifting from "whose model is the biggest" to "whose model is the most cost-effective."

Key Points

• Event time: May 28, 2026 (Liquid AI released) & May 26 (Hy3 list analysis)
• Core changes: Edge/low-cost models enter the mainstream, MoE architecture + small activation parameters become the optimal cost-effective solution
• Direct impact on AI Agent users: The threshold for running high-quality models locally has been greatly reduced, and API costs continue to decline.

Event 1: Liquid AI LFM2.5-8B-A1B — the real “edge AI

On May 28, Liquid AI released the LFM2.5-8B-A1B, a MoE (Mixed Expert) model designed for consumer-grade hardware. List of core parameters:

Most impressive is the leap in benchmark testing. The AA-Omniscience index improved from -78.42 to -24.70 (+53.62), and the non-hallucination rate soared from 7.46 to 63.47 (+56.01). This is a qualitative breakthrough for a model with 8B parameters and only 1B activations.

In terms of tool calling capabilities, BFCLv3 improved from 45.07 to 64.36 (+19.29), and Multi-IF improved from 58.54 to 79.93 (+21.39). This route directly targets AI Agent scenarios - performing complex toolchain calls on the local device.

Liquid AI emphasizes that the positioning of this model is "every token is cheap": the MoE model is naturally suitable for computing constrained environments, and a small number of activation parameters makes the cost of each inference token extremely low. Models already support llama.cpp, MLX, vLLM, SGLang, meaning they can run on M4 Macs, laptops and even Raspberry Pis.

Event 2: Hy3’s mysterious summit — users voted with their feet

At almost the same time, Max Woolf revealed an interesting phenomenon in his blog: a model named Hy3 preview surpassed Claude** on the OpenRouter rankings, becoming the most used model with a gap of more than 50%.

Hy3 is a large language model open sourced by Tencent. The Hugging Face page is simple and the benchmark results are mediocre (comparable to the same level of Chinese open source models), but its usage on OpenRouter is unparalleled. What about the price? $0.066/million tokens – 34% cheaper than DeepSeek V4 Flash ($0.10).

What's even weirder is that the input-output ratio of Hy3 preview is as high as 98% input / 2% output - which means that users almost only use it for "reading" tasks (document processing, information extraction, code review), rather than "writing" tasks (generation, creation). Its use in non-programming scenarios is equally strong.

And this isn't inflated data due to the free trial: Hy3 initially had a free phase (until May 6th), but current ranking data comes from paying users.

Common trend: The market is voting with its feet

These two stories may seem independent, but they tell the same story - the AI market is shifting from supply-driven to demand-driven.

The focus of the AI industry in 2025 is "whose model is the smartest", but by mid-2026, the focus has turned to "who can complete the task at the lowest cost." The ability improvement of large models has entered a plateau period, and users' willingness to pay has become more and more sensitive in the ongoing "sticker shock".

OpenRouter's data is particularly convincing: Hy3 has no large-scale marketing, no word-of-mouth recommendations from the AI ​​community, and even the model quality is not top-notch, but users just choose it because it is cheap enough and easy to use.

Practical inspiration for AI Agent users

1. Local reasoning is no longer a toy: LFM2.5-8B-A1B’s IFEval 91.84 and BFCLv3 64.36 illustrate that the model with 1B activation parameters is already capable of most tool calling tasks. If your agent workflow can run locally, this significantly reduces running costs and latency.
2. The trade-off between cost and quality is being redefined: If you are using Claude or GPT for batch document processing and code review, try using Hy3 or DeepSeek V4 Flash instead - you may find that 80% of the tasks are completed equally, but at only 1/4 the cost.
3. Follow the OpenRouter rankings: It is the most authentic market signal - the results of users voting with real money, which can better reflect "what model is easy to use in actual work" than any benchmark test.

Reference sources

• Liquid AI: LFM2.5-8B-A1B 官方发布
• Max Woolf: Hy3 LLM 登顶 OpenRouter 排行榜分析
• HN Discussion: Liquid AI / Hy3

Tool entry

This article covers the following AI tools and platforms: OpenAI, Claude, DeepSeek, vLLM, Hugging Face, OpenRouter, llama.cpp, MLX

Internal link guidance

• Want to run the model locally? Watch the tutorial: 如何用 LM Studio 在 M4 Mac 上运行本地 AI 模型
• Real cases of making money with low-cost models: 用 DeepSeek V4 + Claude Code 搭微 SaaS 矩阵，月入 8,500 美元
• Want to learn more about low-cost models in action? Watch: DeepSeek Reasonix 实战：零成本搭建 AI 编程 Agent