Axionomy and BSC AI Factory: Open-Weight Agent Models on MareNostrum 5

Axionomy is an AI agent workspace in production for innovation leaders—Think, Run, Watch, and Produce on one harness with RAG, knowledge graph memory, 35+ integrations, and mission observability. Today we announce a strategic cooperation with BSC AI Factory at the Barcelona Supercomputing Center (BSC) to explore training and adapting open-weight language models for specific agent workloads, using world-class supercomputing resources including MareNostrum 5.

Why this partnership matters now

Our platform already orchestrates autonomous agents: planning, tool selection, function calling, contextual memory, error recovery, and deliverables in Workspace. LLM inference today runs primarily through external APIs (including Groq for open-weight models). That model works—but as usage scales, teams feel three pressures: recurring inference cost, latency in multi-step missions, and limited control over which model runs which task inside the harness.

The BSC AI Factory engagement is designed to answer a concrete question: can smaller, specialized open-weight models—roughly in the 20B–30B parameter class—handle targeted parts of the agent loop with acceptable quality, lower cost per task, and latency suitable for interactive innovation work?

What Axionomy brings to the pilot

Axionomy is a SaaS platform with approximately 150 users in production today, an internal API, agent orchestration, tools, RAG over documents and connected services, Neo4j and PostgreSQL/pgvector storage, and strict user-level data separation with authentication and access controls. We own the harness; we want evidence-based options for owning more of the inference stack when it makes sense.

Agent workloads we will study include: interpreting user goals, multi-step planning, tool and function calling, retrieval-augmented generation, long-context memory use, reasoning steps, retries after tool failures, and generation of structured deliverables. The architecture is provider-agnostic—we can swap models per task—but the pilot will establish which tasks merit dedicated fine-tuned models versus general-purpose baselines.

What BSC AI Factory provides

BSC AI Factory supports organizations that need expert guidance on AI and HPC infrastructure, onboarding to supercomputing environments, fine-tuning strategy (supervised fine-tuning, preference optimization, distillation, and related techniques), benchmarking methodology, and trustworthy AI practices (fairness, robustness, privacy, explainability, and transparency).

For Axionomy, that means access pathways to BSC computing resources suited to experimentation, reproducible benchmarks against our production-style agent tasks, and collaboration on datasets and evaluation criteria—not generic leaderboard scores, but task completion, tool accuracy, latency percentiles, and cost per mission.

MareNostrum 5 and the HPC path

MareNostrum 5 is among Europe’s leading supercomputers, operated by BSC. For an AI agent platform, HPC is not academic vanity—it is the practical place to train and stress-test open-weight models before we commit to self-managed inference clusters. The pilot explores whether experiments at BSC can de-risk a later inference infrastructure operated by Axionomy (and potentially offered to partners) once models and operating procedures are proven.

Model candidates and methodology

We will use Groq-served open-weight models as a production baseline and compare against models adapted through the pilot. Initial candidates discussed in our technical cooperation include GPT-OSS 20B, Qwen-class models around 27B parameters (exact Hugging Face identifiers to be confirmed with BSC), and GPT-OSS 120B as a larger reference—not necessarily for every agent step, but to bound quality expectations.

The hypothesis: a smaller model specialized through SFT, preference optimization, or distillation may match or exceed general models on narrow agent skills (for example tool routing or structured planning) while using fewer tokens and milliseconds per step.

We have not yet run a formal in-house fine-tuning program at scale; this pilot is how we build that capability with BSC mentorship. Technical leadership for implementation is led by Ignacio Souto, with an estimated one to two days per week on the pilot, increasing during benchmark sprints.

How we will measure success

Dimensioning assumptions for the pilot (to be replaced by Axionomy production logs) include on the order of thousands of agent tasks per month and multiple LLM calls per task, with heterogeneous workloads—some lightweight, some heavy on RAG, memory, and tools. We will track tokens per call and per task, time-to-first-token, end-to-end task latency (p50/p95/p99), tool-call success and retry rates, and total cost per completed mission.

Design targets for interactive agent UX include roughly sub–2–3 second p95 per LLM call where feasible and complete medium-complexity workflows in roughly 10–15 seconds p95 at the mission level—these are engineering goals for the pilot, not guaranteed production SLAs on day one.

Trust, data, and what we are not claiming

Axionomy’s architecture restricts routine access to user content; operational metadata and anonymized evaluation sets for benchmarking will be defined explicitly with BSC. This announcement describes a cooperation and service proposal (BSC references OPD-2026-01865 and SP-2026-01921)—not finished custom weights shipping to every tenant tomorrow.

If experiments succeed, the intent is to integrate suitable models into inference infrastructure we operate, improving the Axionomy harness for innovation and R&D teams and optionally enabling related services for third parties. Final deployment depends on pilot results and operational readiness.

What this means for Axionomy customers

Near term, nothing breaks: Think (Marita AI), Run (Autonomous Agent), Watch (Mission Observatory), and Produce (Workbench) continue on current providers. Medium term, specialized models could route automatically—cheaper steps on small models, frontier steps only when the mission requires it—without customers re-architecting their programs.

For procurement and research leaders, the headline is stability plus a credible path to sovereign, cost-aware inference aligned with an agent platform you already connect to Gmail, Drive, Notion, LinkedIn, and CRM—not a separate research toy.

What to watch next

We will share benchmark summaries and integration milestones on this blog as the pilot progresses. Explore harness modes in our docs, read AI Agent Features in the Wild for day-to-day use cases, or start a workspace at https://www.axionomy.xyz while inference experiments run in parallel.

We thank BSC AI Factory for partnering with Axionomy on this next chapter—specialized models, supercomputing rigor, and agent outcomes that innovation leaders can trust and measure.