PeopleofSuperintelligence

Builder

Mathis Felardos

Also known as hasB4K

Mathis Felardos is an inference software engineer at Mistral AI whose work connects open models with the systems that serve them. In 2025, he coauthored reports on Magistral's reasoning pipeline, Voxtral audio chat and Devstral-Small for code agents. He presented Mistral's disaggregated vLLM serving work in Zürich in November 2025 and published a vLLM memory-leak investigation in January 2026. In March 2026, he described Mistral's contribution to llm-d, including development of the DisaggregatedSet operator for LeaderWorkerSet; in September 2026, he joined Kubernetes SIG's LeaderWorkerSet approvers. As of October 2026, he works as an inference engineer at Mistral AI in Paris and contributes to llm-d.

Timeline

  1. June 2025

    Magistral reasoning report

    In June 2025, he coauthors the Magistral report, submitted to arXiv in June 2025, which introduces Mistral's first reasoning model and its scalable reinforcement-learning pipeline. It presents Magistral Medium, trained for reasoning on top of Mistral Medium 3 with RL alone, and open-source Magistral Small.

  2. July 2025

    Voxtral audio models

    In July 2025, he coauthors the Voxtral report, which presents Voxtral Mini and Voxtral Small as multimodal audio chat models released under Apache 2.0.

  3. August 2025

    Devstral for code agents

    In August 2025, he coauthors the Devstral technical report. It introduces Devstral-Small, a lightweight 24B open-source model for code agents.

  4. November 2025

    Disaggregated serving in Zürich

    In November 2025, he and Mistral GPU engineer Mickaël Seznec present “P/D Disaggregated Serving in Production with vLLM” at the first official vLLM meetup in Europe in Zürich.

  5. January 2026

    Debugging a vLLM leak

    In January 2026, he publishes “Heaps do lie: debugging a memory leak in vLLM,” opening Mistral's Engineering Deep Dive series. The team traces a leak in disaggregated serving of Mistral Medium 3.1 to the UCX communication library and merges a fix into vLLM.

  6. March 2026

    DisaggregatedSet for LeaderWorkerSet

    In March 2026, llm-d joins the CNCF as a Sandbox project. He describes Mistral's contribution to llm-d, including development of a DisaggregatedSet operator for LeaderWorkerSet.

  7. September 2026

    LeaderWorkerSet approver

    He joins the OWNERS approvers for Kubernetes SIG's LeaderWorkerSet project in September 2026.

  8. As of October 2026

    Inference engineer at Mistral

    As of October 2026, he works as an inference software engineer at Mistral AI in Paris and contributes to llm-d. His engineering work has included vLLM, a high-throughput and memory-efficient inference and serving engine for LLMs.

  1. The median h-index is 48 at universities and 14 at frontier AI labsIncludes Mistral AI. The median h-index is 48 at universities and 14 at frontier AI labs. We compare publishing careers, recent citations, dual appointments and job moves across 2,420 university researchers and 308 frontier-lab researchers.
  2. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Mistral AI. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.
  3. Of 943 people shaping AI who worked at Google, 446 work there nowIncludes Mistral AI. Our estimates of where the 943 people shaping AI who worked at Google and Google DeepMind work now: 446 at Google and 452 across 278 other organizations.
  4. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staffIncludes Mistral AI. 5,000 frontier users would be 38.2% of the core labs' 13,100 reported staff. We test the estimate against 3,913 named model contributors and dated all-staff records across Google DeepMind, OpenAI and Anthropic, plus six additional frontier-grade labs.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Mathis Felardos yet. Send us a free-licensed photo.

  1. [2506.10910] Magistral arxiv.org. Checked 21 October 2026.
  2. [2507.13264] Voxtral arxiv.org. Checked 21 October 2026.
  3. [2509.25193] Devstral: Fine-tuning Language Models for Coding Agent Applications arxiv.org. Checked 21 October 2026.
  4. Zürich vLLM Meetup · Luma luma.com. Checked 21 October 2026.
  5. Heaps do lie: debugging a memory leak in vLLM. | Mistral AI mistral.ai. Checked 21 October 2026.
  6. Welcome llm-d to the CNCF: Evolving Kubernetes into SOTA AI infrastructure | CNCF cncf.io. Checked 21 October 2026.
  7. Add hasB4K to OWNERS approvers by hasB4K · Pull Request #1077 · kubernetes-sigs/lws · GitHub github.com. Checked 21 October 2026.
  8. llm-d 1.0, an open source stack for distributed inference | Google Cloud Blog cloud.google.com. Checked 21 October 2026.
  9. hasB4K (Mathis Felardos) · GitHub github.com. Checked 21 October 2026.