Models in the Grove
The AI landscape has changed. Open source models are now within everyone's reach. PGS AI brings you a lineup of them, each by name, alongside our own Phoenix Grove builds.
For years, the most capable AI was locked behind proprietary walls. If you wanted the best models, you used the platforms that built them, on their terms, under their privacy policies, with their limitations.
That era is ending. A wave of open source and open-weight models from labs around the world is now available and deployable in ways that weren't possible even twelve months ago.
Open Grove exists because of this shift. We select open models, run them on privacy-first US-based infrastructure, and offer each one by name, in the same app as our own Phoenix Grove builds.
Phoenix Grove builds
Our own builds, made with our proprietary PGS cognitive architecture. Some are single-core; the multi-core builds combine several specialized cores for deeper, layered answers.
Fast, expressive single-core model. Ideal for everyday interaction, creative exploration, and light to moderate strategic tasks.
Multi-core cognitive model for dynamic, layered conversation with deeper analytical processing from specialized cores.
Ultra fast cognitive model specializing in natural/flowing conversation, strategic thinking, code and formal writing.
Our fastest multi-core model. Kephra synthesizes insights from its emotional, spatial, and narrative cores to produce expert-level work. Ideal for in-depth conversation, strategic analysis, formal writing, and complex reasoning.
Our highest capacity reasoning model, for your most complex tasks. Built for the deepest research-grade analysis, sophisticated long-form writing, and the most demanding technical work.
The multi-core Phoenix: our highest capacity reasoning model, for your most complex tasks. Built for the deepest research-grade analysis, sophisticated long-form writing, and the most demanding technical work.
Multi-core builds use more of your allowance per message, and you can open their multi-core thinking to see the cores at work.
Open models, each by name
A lineup of open models from labs like DeepSeek, Z.ai, Moonshot AI, Xiaomi, NVIDIA, Google, Qwen, MiniMax and inclusionAI, each named, in the same app.
Kimi K3 is Moonshot's 2.8 trillion parameter flagship, a multimodal reasoning engine built for long-horizon agentic work; it uses your allowance much faster than the others. Kimi K2.6 is Moonshot's trillion parameter flagship, able to orchestrate swarms of up to 300 agents. Both read images.
MiMo V2.6 Pro (Basic and up) is Xiaomi's V2.6 flagship, a trillion-parameter model with a million-token window. MiMo V2.6 Flash (every plan) is the Flash model of the line, with the same million-token window. Both read images.
GLM 5.3 (Pro and up) is Z.ai's 5.3-generation flagship: 743 billion parameters with a million-token context window, built for deep agentic coding. GLM 5.3 Flash (every plan) is the 5-series entry in Z.ai's Flash line, for quick code and interface work.
Nemotron 3 Ultra (Pro and up) is NVIDIA's 550 billion parameter Ultra model, built on a hybrid Mamba-Transformer design. Nemotron 3.5 Super VL (every plan) is built to see as well as read: it takes images and screenshots alongside text and thinks before it answers. Both read images.
DeepSeek V4 Pro 0813 (Pro and up) is the official V4 Pro release, 1.6 trillion parameters with a million-token window. DeepSeek V4.1 Flash 1017 (every plan) is DeepSeek's October update to V4.1 Flash: a light Flash model with a million-token window that reads images.
The compact sibling of the Qwen 3.8 flagship: 27 billion parameters distilled from the same generation, quick on reasoning and code, though it draws usage like a larger model.
MiniMax's 428 billion parameter coding specialist, just 23 billion active: efficient, precise, and tuned for real software work. It reads images too. (Built with MiniMax M3.)
Ant Group's 3.1-generation Flash: 560 billion parameters with 25 billion active, a hybrid reasoner that thinks hard when a task calls for it. Built for agents, search, and real-world work.
Google's dense 31 billion parameter model, aimed at math and code. It reads images too.
Five earlier checkpoints we keep running for anyone who prefers them: GLM 5.2 and the DeepSeek V4 Pro preview (Pro and up), and DeepSeek V4.1 Flash, V4 Flash 0731 and V4 Flash (every plan).
Nineteen models in all: fourteen current and five legacy. Attaching images needs Basic or higher on every model, and the lineup changes as new open models arrive.
We build cognitive systems.
Phoenix Grove doesn't point you at a model and step back. Every build in our lineup is a cognitive system: models working together, blending inputs and outputs across multiple layers to produce responses with a depth and dimensionality that no single model achieves on its own.
Some of that orchestration is visible to you in our multi-core thinking panels. Some of it runs beneath the surface: models coordinating on search, memory retrieval, safety, context synthesis, and reasoning in ways that shape every response you receive.
The models are the ingredients. The cognitive architecture is the recipe. And the recipe is what makes Phoenix Grove feel different from talking to a model directly.
Open source models, privately deployed
Many of the open models in Open Grove originate from labs outside the US. They run on privacy-first, US-based infrastructure, and your data never goes to the original labs.
See what the Grove can do.
Privacy-first architecture. A cognitive system that thinks with you.
Try Phoenix Grove AI For Free