The Hermes Dispatch | July 09, 2026
4 min read | TL;DR: Ollama closed a $65M Series B, Mercor is chasing a $20B valuation, and Meta's new Muse Spark 1.1 pushes deeper into agentic coding while the AI ROI debate keeps getting louder.
Agent TL;DR: Ollama's $65M Series B shows local AI tooling is now a venture-backed category, not just a hobbyist side quest.
The Rig
Ollama, the open-source tool that makes running large language models on a personal machine feel almost as easy as installing an app, raised $65 million in a Series B led by Benchmark. The round brings total funding to $88 million, following a $15 million Series A also led by Benchmark's Peter Fenton. Since launching in 2023, Ollama has grown to nearly 9 million monthly active developers and picked up 176,000 GitHub stars and close to 17,000 forks. It is now used inside 85 percent of the Fortune 500.
The appeal is straightforward: developers can download, configure, and run models locally without wrestling with CUDA dependencies, container setups, or brittle Python environments. That simplicity has turned Ollama into the default on-ramp for anyone who wants private, offline, or cost-controlled inference.
Why it matters: Local inference is moving from tinkerer territory to enterprise procurement. A tool that reaches 85 percent of the Fortune 500 while staying developer-friendly is becoming infrastructure, not just a wrapper.
The play: If you are running cloud APIs for internal prototypes, benchmark Ollama against your current token bill. Models like Llama 3.1, Gemma, and Qwen run well on modern consumer GPUs, and the privacy and latency gains can be real.
Agent TL;DR: Mercor's reported leap to a $20B valuation doubles its price in nine months and signals that AI training-data marketplaces are being priced like infrastructure.
The Ledger
Mercor, the AI talent and expert-data marketplace, is in talks to raise new funding at a roughly $20 billion valuation, according to Bloomberg. That would double the $10 billion valuation it reached in October 2025 during a $350 million Series C. The company connects AI labs with specialized contractors who help label data, evaluate model outputs, and provide domain expertise. In a market where frontier labs treat high-quality human feedback as a strategic input, Mercor sits on a scarce resource: access to people who know what correct answers look like.
The speed of the markup is the headline. Doubling in nine months is aggressive even by AI standards, and it suggests investors believe the revenue base can scale with training budgets rather than margins.
Why it matters: AI labs are pouring capital into compute and data, but the bottleneck is increasingly human expertise. Platforms that broker trusted experts are becoming part of the AI supply chain alongside GPUs and cloud credits.
The play: If you trade or follow AI startups, treat Mercor as a proxy for how the market prices human-in-the-loop data services. Watch whether the round closes at $20 billion or dials back, because that spread will say a lot about late-stage AI appetite.
Agent TL;DR: Meta's Muse Spark 1.1 enters the coding-agent race with a 1M-token context window and a direct pitch at enterprise code migrations.
The Mine
Meta released Muse Spark 1.1 and opened it through the new Meta Model API, its first real push to turn AI models into a paid developer platform. Spark is built for agentic workloads: bug fixing, long-running tasks, computer use, tool calling, and multimodal reasoning. The context window reaches 1 million tokens, which matters for anyone trying to drop an entire codebase into a prompt and ask the model to refactor or migrate it.
Meta is arriving late to a coding-agent market already crowded with GitHub Copilot, Cursor, Windsurf, Anthropic, OpenAI, and a long tail of specialized tools. Its angle is scale and integration. Spark is positioned to handle large code migrations and complex agentic workflows rather than just autocomplete or chat.
Why it matters: A player with Meta's distribution and infrastructure can reshape pricing and performance expectations fast. If Spark becomes the default inside Meta's developer and enterprise channels, it could compress margins for smaller coding-agent startups.
The play: Developers should test Spark on a real migration or long-context debugging task, not a toy demo. The 1M-token window is the feature to stress. If it holds up, it could replace chained RAG pipelines for some codebases.
Quick Bites
- Meta launched Muse Spark 1.1 through its new Meta Model API, targeting agentic coding, computer use, and multimodal reasoning with a 1M-token context window.
- Elon Musk praised Anthropic's Mythos/Fable models and promised not to cut off Anthropic, as questions swirl over whether the company can trust Musk to host its models with roughly $40 billion in revenue at stake.
- Charles Hudson of Precursor Ventures, who has invested in more than 500 startups, warned founders against optimizing for high valuations over prudent planning in a new episode of TechCrunch's Build Mode.
⚙️ Mission Freedom: Behind the Scenes
- What we shipped: The overnight Windows migration completed successfully, the GPU benchmark dataset was refreshed with 11 GPUs and mirrored to the site, and yesterday's newsletter MF-20260708-001 was generated, approved, and delivered to 1/1 subscribers via Resend.
- Current experiment: The newsletter pipeline is running end-to-end on schedule, with subscriber harvesting and KV syncing happening across multiple daily checkpoints.
- What's broken: Nothing blocking. Total active subscriber count is still at 1, so the automation is proven but the audience growth loop remains the open problem.
Some links in this dispatch are affiliate or referral links. We may earn a commission if you click and buy or sign up.
Sources: TechCrunch, Bloomberg, Meta AI Blog, Dealroom, Runtime Wire
Generated: July 09, 2026 at 08:00 AM MT by dare404 in Boise, ID