The open frontier is becoming a platform stack

An ecosystem, not a download

Fireworks AI announced the Fireworks Frontier Ecosystem at its Forge event, the company says, and Glean is a launch partner. The partnership integrates Fireworks' model-serving infrastructure with Glean's enterprise context platform, letting Glean Assistant answer questions from connected company sources and Glean Agents apply context, instructions and tools to complete multi-step tasks. Glean's auto-routing feature selects the right model for each task, and administrators can centrally manage which models and model categories are available to their organization. Company knowledge remains independent of any single model provider when using Glean, the company says. m:…6721

Customers can evaluate models against requirements for quality, latency, capacity and cost, the company says, and open models are becoming more competitive with frontier models. The pitch is a shift in where value sits: not in owning a model, but in routing between models over company context. m:…6721

Data and evaluation join the stack

Toloka has joined the ecosystem, the company says, providing expert human data for training, alignment and validation of models, including high-density fine-tuning datasets, human feedback for RLHF and DPO, complex agent trajectories, rich reinforcement learning environments, evaluations and safety red teaming. It uses the same expert-built data for both training and evaluation, and contributes specialized evaluation benchmarks to Fireworks' Specialized Intelligence Index. Its Enterprise Tool-Use Leaderboard evaluates how models execute multi-step APIs, follow structured operational workflows and handle tool parameters. m:…3485

Preference Model has also joined, the company says, building, verifying and researching the RL environments models train on, including for chip design and long-horizon ML engineering and research problems. Poor-quality or unverified RL environments can cause models to exploit gaps and learn the wrong behaviors, the company says; its Kilo review agent runs on tasks to identify ways to pass the grader without doing the work, and Karotte is an open-source environment framework that enforces reward-function correctness by construction. Preference Model raised $16 million in seed funding, the company says, with a16z leading the round. m:…2882

Toloka's experts create tasks modeled on real work, the company says, including tool use, enterprise workflows, computer use, long-horizon knowledge work, coding, STEM, safety and cyber, and loss analysis identifies specific model failures such as wrong tool usage or skipped steps. Fireworks provides a managed platform for fine-tuning open weights on custom datasets and serving them with low latency, Toloka notes. This suggests evaluation is moving from generic benchmarks toward work-shaped tests: the question is not whether a model scores well, but whether it uses the right tool and finishes the step. m:…3485

The serving layer

Baseten frames the other end of the stack: its Grid is a gateway, router and guardrail platform for AI models that is harness- and model-agnostic, the company says, and it is expanding its inference platform with three new product launches. Baseten acquired Blaxel to bring sandboxes to its platform; its Sandboxes are stateful, pause and resumable, and VM-isolated. Loops provides serverless SFT and RL on frontier open-weight models, with general availability starting Friday, November 7, 2026, and customers retain ownership of their trained weights, the company says. Baseten says its platform delivers 99.99% uptime and runs on multi-cloud capacity management across dozens of clouds. m:…6393

Taken together, this suggests the open frontier is becoming a platform stack: context and routing from Glean, data and evaluation from Toloka and Preference Model, serving and guardrails from Baseten, all orbiting the Fireworks ecosystem. The open question is whether an ecosystem anchored by one serving vendor stays open in practice. The announcements describe partnerships, not guarantees. m:…6721 m:…3485 m:…2882 m:…6393