InfiX.ai

助力生成式人工智能的
最後一公里

我們的使命與願景

InfiX.ai 正在通過低精度訓練、模型融合、GUI 智能體和多模態智能，開創協作式生成 AI 的新路徑

使命

在 InfiX.ai，我們致力於讓 AI 更易獲取、更高效、更強大。我們的研究專注於開發創新技術，在保持實際可用性的同時，不斷突破 AI 系統所能實現的邊界。

願景

我們不只是在構建又一個模型，而是正在構建能夠構建所有模型的系統。

研究重點

推理增強型低精度訓練

降低訓練時間、內存壓力和推理成本，提升模型訓練與部署效率。在保持接近 BF16 精度的同時，實現最高 22% 的訓練時間縮短和 19% 的吞吐量提升。

模型融合與模型合併

將多個模型的優勢融合為更強的專家級 AI，同時無需承擔多模型推理的額外開銷。支持跨多領域的百億參數級基礎模型融合，所需 GPU 算力僅為完整預訓練的萬分之一。

推理增強型低精度訓練

降低訓練時間、內存壓力和推理成本，提升模型訓練與部署效率。在保持接近 BF16 精度的同時，實現最高 22% 的訓練時間縮短和 19% 的吞吐量提升。

模型融合與模型合併

將多個模型的優勢融合為更強的專家級 AI，同時無需承擔多模型推理的額外開銷。支持跨多領域的百億參數級基礎模型融合，所需 GPU 算力僅為完整預訓練的萬分之一。

最新研究

2026年4月30日

BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models

Large language model inference is often bounded by memory footprint and memory bandwidth in resource-constrained deployments, making quantization a fundamental technique for efficient serving. While post-training quantization maintains high fidelity at 4-bit, it deteriorates at 2-3 bits. BPDQ constructs a variable quantization grid via bit-planes and scalar coefficients, then iteratively refines them with approximate second-order information and progressive error compensation to minimize output discrepancy. In the 2-bit regime, BPDQ enables serving Qwen2.5-72B on a single RTX 3090 with 83.85% GSM8K accuracy, compared with 90.83% at 16-bit.

2026年4月30日

InfiMed-ORBIT: Aligning LLMs on Open-Ended Complex Tasks via Rubric-Based Incremental Training

ORBIT is an open-ended rubric-based incremental training framework for high-stakes medical dialogue. It combines synthetic dialogue generation, retrieval-augmented rubric construction, difficulty filtering, and rubric-guided reinforcement learning to align LLMs on context-dependent consultation tasks. Applied to Qwen3-4B-Instruct, ORBIT raises the HealthBench-Hard score from 7.0 to 27.5 using only 2k training samples, with larger rubric sets improving performance to 33.6 and 37.3.

2026年4月30日

Model Merging Scaling Laws in Large Language Models

This work studies empirical scaling laws for language model merging measured by cross-entropy. Across 10,866 merged models, base sizes from 0.5B to 72B, nine domains, and four methods, the paper identifies a compact floor-plus-tail power law that links model size and expert count. The law explains why most merging gains arrive early, why variability shrinks as more experts are included, and how to plan expert count under a budget.

查看所有研究