FriendliAI Expands High-Performance AI Model Offering With Gemma-4-31B-it

According to a recent LinkedIn post from FriendliAI, the company is highlighting the availability of Google DeepMind’s Gemma‑4‑31B‑it model on its Model APIs and Dedicated Endpoints. The post describes Gemma‑4‑31B‑it as a dense, instruction‑tuned, multimodal vision‑language model positioned as the largest open‑weight model in the Gemma 4 family.

The company’s LinkedIn post suggests that its Gemma‑4‑31B‑it deployment leads third‑party Artificial Analysis rankings for output speed, time‑to‑first‑token, and overall response times. FriendliAI also emphasizes configurable “thinking” and native function calling, targeting use cases such as coding, agentic workflows, document extraction, and question answering.

Benchmark figures cited in the post include a reported 2,150 Codeforces ELO and 80.0% on LiveCodeBench v6 for agentic coding performance. Additional metrics include 89.2% on AIME 2026 and 84.3% on GPQA Diamond for math and reasoning, as well as 76.9% on MMMU Pro and 85.6% on MATH‑Vision for document understanding and multimodal tasks.

The LinkedIn post indicates that Gemma‑4‑31B‑it can be accessed in a serverless configuration through FriendliAI’s Model APIs or deployed with a single click on Dedicated Endpoints. For investors, this positioning may underline FriendliAI’s focus on high‑performance inference infrastructure and its efforts to differentiate on latency and capability in the competitive market for hosting large open‑weight AI models.

If the performance claims and benchmark scores referenced in the post are sustained in production workloads, FriendliAI could strengthen its value proposition for developer and enterprise customers seeking cost‑efficient, low‑latency model serving. This may support usage growth, deepen integration with client applications, and potentially enhance the company’s standing as an infrastructure partner within the rapidly evolving generative AI ecosystem.