The Harness

Local models clear the daily-coding bar

Show Notes

Today's 921-point HN thread confirms local models have crossed the daily-coding threshold: Qwen 3.6 MoE is the consensus pick, economics now favor on-prem routing for constrained tasks, and the remaining gaps are harness problems, not capability ceilings. Microsoft shipped Work IQ API to GA today, turning M365 data into the competitive moat for enterprise agents billed on Copilot Credits. OpenAI's audited $34B 2025 spend and the DOJ's national security designation for xAI both set the terms for AI's next phase.

What is The Harness ?

A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.

Good morning, it's Tuesday, June sixteenth.

In today's briefing we see local models crossing the daily-coding bar on consumer hardware, Microsoft shipping Work IQ to unlock enterprise agent deployments through M365 data access, and OpenAI's audited 2025 spend clarifying the financial stakes of frontier AI.

First up - Today in the big model news;

Open AI

OpenAI's audited financials for two thousand twenty-five are now public: thirty-four billion dollars in total expenditure, nineteen billion on research and development, and nearly six billion on sales and marketing. These numbers land as the company works through IPO review with Goldman Sachs and Morgan Stanley. The gap between this burn rate and any plausible two thousand twenty-five revenue is the central question for public investors. It also explains why Codex is getting aggressive distribution right now: free Pro subscriptions for open-source maintainers, AWS Bedrock placement, and mobile rollout. For AI PMs shipping coding-centric products, this audited burn rate and Codex distribution signal a market shift where coding agents become the primary growth lever, because the financial gap between expenditure and revenue forces scaling across all available channels.

Local model developments

A nine hundred twenty-one upvote Hacker News thread gives today's real-time community verdict: local models have crossed the daily-coding bar on consumer hardware. Qwen three point six, thirty-five B A three B, a mixture of experts model with three billion active parameters, is the consensus pick. On a Mac Studio with one hundred twenty-eight gigabytes of RAM or a dual RTX 3090 setup, it runs at interactive speeds of one hundred to one hundred fifty tokens per second at three hundred K context length. MiniMax M three posts fifty-nine percent on SWE-Bench Pro as the first open-weight model combining frontier coding capability, one million token context, and native multimodality. The documented failure modes are consistent: edit tool imprecision, reasoning overhead, and cache management issues. These are harness problems, not fundamental capability limitations. For AI PMs thinking about per-engineer cost economics and task routing, there is now a credible spreadsheet case for bifurcating work between local and frontier models, because what runs on consumer-grade machines has crossed the threshold where it handles real production coding work reliably.

In the harness, tools and orchestration world;

Microsoft shipped Work IQ to general availability today, grounding enterprise agents in M365 data at scale. The service taps email, calendar, meetings, files, people, and collaboration patterns across three hundred million Office users through an A2A protocol, a redesigned remote MCP server, and a REST API. Billing converts to Copilot Credits, the same consumption currency as Copilot Studio, eliminating the per-user license overhead that blocked earlier pilots. For product teams building knowledge-worker agents, expect enterprise procurement to bifurcate sharply between Microsoft-integrated stacks and competing harnesses, because any alternative agent system that doesn't tap Work IQ will struggle to win Fortune five hundred automation deals against Microsoft's Copilot.

In other news;

The DOJ declared xAI integral to US military operations in a filing seeking dismissal of an NAACP lawsuit. This is the first formal government assertion of Grok's military role. The designation complicates civil rights challenges, creates regulatory shelter beyond what OpenAI or Anthropic currently enjoy, and positions xAI for defense procurement in ways that may bypass standard competitive contracting. For companies evaluating national security and government procurement dynamics in AI, xAI's military designation creates a different competitive baseline, because the DOJ filing grants Grok strategic-asset status that opens procurement pathways neither OpenAI nor Anthropic currently enjoy.

That's the briefing. Have a great day.