Shanda founder's 27B local model takes the agent fight to DeepSeek
StartLux's first preview beats much bigger models on agent benchmarks, VAST ships production-ready 3D scene generation from a single image, and Hugging Face gives coding agents a memory the user controls.
Shanda co-founder Chen Danian's StartLux unveiled a 27-billion-parameter local agent model on Wednesday that, by its own numbers, runs within striking distance of DeepSeek's trillion-parameter flagship on tool-using tasks. The StartLux-V1.0-27B-Preview outperforms the 284B DeepSeek-V4-Flash and the 198B Step-3.7-Flash by 1.3 percentage points on a composite agent benchmark that scores navigation, web search, browser automation, financial analysis, code-repo management and 3D design, and beats the same-sized Qwen 3.6 by more than five points. In a controlled browser task — finding the cheapest direct economy flight from Singapore to Beijing on a specific date — StartLux picked a 299-dollar ticket in about 95 seconds; Claude Sonnet 4.6 returned a 556-dollar option after more than twice the runtime. The bet is that on-device models can collapse the cloud-cost curve for long agent tasks: instead of accumulating API tokens, the device pays for power and silicon once.
The numbers matter less than the framing. Chen, the co-founder of Shanda Network and the WiFi-master-key operator LinkSure, is not pitching another trillion-parameter flagship. He is arguing that post-training on agent traces — choosing tools, recovering from errors, knowing when to declare a task done — buys more capability per parameter than raw scale ever did. StartLux plans to train its own base model next, with diffusion-style architectures on the roadmap. The thesis lands at a moment when the industry is quietly conceding that parameter count is no longer the variable that decides agent quality.
VAST's WorldGen turns a single image into a fully editable 3D scene, the first 3D system to reach production-grade scene-level generation. Most image-to-3D models collapse an entire scene into one un-editable blob; if you want to move a chair, you start over. WorldGen splits the output into Lego-like modules — every object is its own asset with its own physics, so you can pick up the bowl, swap the chair, or remove a load-bearing prop and watch the surrounding objects react. The system is built on CAST, a SIGGRAPH Best Paper from last year that spent the year since solving the error-accumulation problem that comes with chaining multiple generative modules. WorldGen is already in closed beta with game studios and film teams, and the team says it is shipping into live productions over the coming months. The argument is that 3D is the AI application closest to production, and the only way to unlock that value is to make every generated piece editable, reusable, and editable downstream.
For game designers this collapses the workflow from a chain of human asset assembly to a starting scene that level designers, artists, and directors can already touch. For film teams, WorldGen acts as a controllable 3D base that video models can paint on top for final shots — Apple Vision Pro and other XR headsets are an obvious endgame. VAST is positioning itself as "the Anthropic of 3D generation" by solving the hard production problem rather than chasing another flashy demo.
Hugging Face launched Funes on Wednesday, an open-source memory layer that lets coding agents carry their context across sessions and across machines — and lets the user own the data. Most agents either compress sessions into summaries that flatten the findings that mattered, or rent memory through a third-party API that locks the user in. Funes skips both: local memory is a Lance dataset, shared memory is a private Hugging Face dataset the user controls, and the agent recalls from it directly inside the conversation rather than asking a remote service. In Hugging Face's own benchmark, recall beat compaction 8x on one task and 4x on the other by cost alone — and unlike compaction, recall returns the actual passages so a finding doesn't have to survive summarization. The same memory can be picked up by Claude Code this week and Codex next week, regardless of which model wrote to it.
The user-ownership angle is the part that matters most. A coding agent's memory is a record of everything you have asked it to do — including the parts you would rather not hand to a vendor. Funes makes that memory portable, private, and versioned through the Hugging Face Hub, the same infrastructure anyone already uses for datasets and models. If agent memory is going to become a real layer of the developer stack, it needs to look a lot more like a dataset than a SaaS subscription. Funes is the first serious attempt to make that the default.
What to watch: whether StartLux's small-model agent bet holds up outside the benchmarks Chen cherry-picked, and whether VAST's production-pipeline partners actually ship scenes built on WorldGen this quarter.
Do you think local models can really take 80% of the cloud agent market in three years, or is StartLux overselling what post-training can buy? Tell us in the comments.
Sources: QbitAI · QbitAI · Hugging Face · Funes on GitHub · Funes memory dataset · Funes handoff-vs-recall benchmark
