Microsoft explores fine-tuned DeepSeek V4 to power Copilot Cowork

Microsoft confirmed it is exploring a fine-tuned version of China's DeepSeek V4 — or another open-source model — to power its Copilot products, a move first reported by Axios and likely to draw political scrutiny in Washington. The driver is economics: agentic tools like Copilot Cowork, Claude Code, and Codex repeatedly call models in long-running loops, a pattern Axios dubbed 'tokenmaxxing' that drives compute bills sharply higher. Microsoft confirmed Copilot Cowork customers will pay based on compute usage, making per-call model cost a direct margin lever.

A cheaper, open-weight, fine-tunable model like DeepSeek V4 (a 1.6-trillion-parameter mixture-of-experts architecture) could undercut the cost of routing everything through OpenAI's or Anthropic's frontier APIs. That logic mirrors the broader open-weights momentum this week — GLM-5.2 topping leaderboards, Hugging Face declaring 'open weights are now our default.'

The geopolitical friction is the catch. Routing a flagship Microsoft enterprise product through a Chinese-developed model — even one fine-tuned and self-hosted — invites scrutiny over data trust and national security, especially as the U.S. weighs blacklisting DeepSeek among 100+ firms flagged as risks. Gizmodo framed it bluntly as a move 'probably to Trump's chagrin.'

This lands the same week Satya Nadella announced Copilot Cowork's worldwide GA with multi-model support (9,600+ likes on LinkedIn) — the multi-model architecture is precisely what makes swapping in DeepSeek feasible. The open question: will Microsoft actually ship a China-origin model in Copilot, or is this leverage to pressure OpenAI on pricing? Watch for whether the exploration survives political pushback.