AMD rolls out open-source OLMo LLM, to compete with AI giants

AMD has launched its first open-source large language models (LLMs) under the OLMo brand, aiming to strengthen its position in the competitive AI landscape led by giants like Nvidia, Intel, and Qualcomm.

AMD OLMo is a series of 1-billion parameter large language models trained from scratch using trillions of tokens on a cluster of AMD Instinct MI250 GPUs. They are designed to excel in reasoning, instruction-following, and chat while embracing an open-source ethos that allows developers access to data, weights, training recipes, and code.

“Continuing AMD tradition of open-sourcing models and code to help the community advance together, we are excited to release our first series of fully open 1 billion parameter language models, AMD OLMo,” AMD said in a statement.

AMD’s open-source approach positions OLMo as an accessible and scalable option for companies seeking alternatives in AI technology. The model can be deployed in data centers or on AMD Ryzen AI PCs equipped with neural processing units (NPUs), allowing developers to leverage advanced AI directly on personal devices, the statement added.

“AMD is following Nvidia’s lead by expanding into the large language model (LLM) space alongside its well-established strength in computing hardware — a direction that Intel and Qualcomm have not yet fully embraced,” said Abhigyan Malik, practice director at Everest Group. “By fostering an open ecosystem, AMD enables developers to innovate and build diverse applications through a network effect.”

According to Malik, this strategy amplifies AMD’s core value proposition, particularly in driving demand for its underlying hardware, including AMD Instinct MI250 GPUs and Ryzen CPUs, where “AMD seeks to create lasting market impact.”

Extensive training and fine-tuning

The OLMo series follows a detailed three-phase training and fine-tuning process, according to AMD.

Initially, OLMo 1B was pre-trained on a subset of the Dolma v1.7 dataset using a transformer model focused on next-token prediction. This helped the model grasp general language patterns. In its second phase, the OLMo 1B was supervised and fine-tuned (SFT) on multiple datasets to refine its capabilities in science, coding, and mathematics.

The final model, OLMo 1B SFT DPO, was optimized with Direct Preference Optimization (DPO) based on human feedback, resulting in a model that effectively aligns its responses with typical user expectations.

Competitive performance and benchmark success

In internal benchmarks, AMD’s OLMo models performed well against similarly sized open-source models, such as TinyLlama-1.1B and OpenELM-1_1B, in multi-task and general reasoning tests, the company claimed. Specifically, its performance increased by over 15% on tasks in GSM8k, a substantial gain attributed to AMD’s multi-phase supervised fine-tuning and Direct Preference Optimization (DPO). ‘

In multi-turn chat tests, AMD claimed, OLMo showed a 3.41% edge in AlpacaEval 2 Win Rate and a 0.97% gain in MT-Bench over its closest open-source competitors.

However, when looking at the broader LLM landscape, Nvidia’s GH200 Grace Hopper Superchip and H100 GPU remain leaders in LLM processing, particularly for large, multi-faceted AI workloads. Nvidia’s focus on innovations like C2C link, which accelerates data transfer between its CPU and GPU, gives it an edge, providing a speed advantage for high-demand inference tasks such as recommendation systems.

Intel, while slightly behind in peak speed, leverages its Habana Gaudi2 accelerator for cost-effective yet robust performance, with future upgrades planned for increased precision. ‘

Meanwhile, Qualcomm’s Cloud AI100 emphasizes power efficiency, meeting the needs of organizations seeking high AI performance without the extensive energy demands associated with Nvidia’s high-end systems.

AMD’s OLMo models also showed strong performance on responsible AI benchmarks, such as ToxiGen (for toxic language detection), crows_pairs (bias assessment), and TruthfulQA-mc2 (accuracy). These scores reflect AMD’s commitment to ethical AI, an essential focus as AI integration scales across industries.

AMD’s position in the AI market

With its first open-source LLM series, AMD is positioned to make significant inroads in the AI industry, offering a compelling balance of capability, openness, and versatility to compete in a market currently led by Nvidia, Intel, and Qualcomm.

However, AMD’s ability to close the gap will depend on how well its open-source initiative and hardware enhancements keep pace with rivals’ advances in performance, efficiency, and specialized AI capabilities.

“AMD’s entry into the open-source LLM space strengthens the ecosystem, potentially lowering the operational costs associated with adopting generative AI,” said Suseel Menon, practice director at Everest Group.

AMD’s move into LLMs places it against established players like Nvidia, Intel, and Qualcomm, who have gained market prominence with their proprietary models.

“This move also puts pressure on proprietary LLMs to continually innovate and justify their pricing structures,” Menon added.

Analysts believe AMD’s unique open-source strategy and accessibility aim to attract enterprises and developers looking for flexible, affordable AI solutions without proprietary constraints.

“For large enterprises with long-term data privacy concerns, AMD’s open-source model offers a compelling alternative as they navigate AI integration,” Menon added. “By building a cohesive, full-stack AI offering that spans hardware, LLMs, and ecosystem tools, AMD is positioning itself with a distinct competitive edge among leading silicon vendors.”