Models

OpenAI's open-weight 21B Mixture-of-Experts model (3.6B active parameters), released under Apache 2.0. Built for low-latency, high-throughput inference, it runs comfortably on a single GPU and handles classification, extraction, routing and short-form summarization well above its size class. That combination has made it one of the…

This provider hasn't added a description for this listing yet. When they do, their summary will show up here alongside pricing and routing details.

Microsoft's small open-weight model. Trained heavily on curated + synthetic data, so it punches above its size on reasoning but knows less about the world. Very cheap to run. Good for tightly-scoped work: classification, routing, structured extraction.

NVIDIA's mid-size open-weight model, tuned for agentic and retrieval-augmented pipelines and optimized to run efficiently on NVIDIA inference stacks. Good throughput and dependable instruction-following at a low price, with solid reasoning for its tier. Widely available across providers, it's a reasonable pick for RAG and tool-use…

Google's open Gemma 3, instruction-tuned at 27B parameters. Multilingual out of the box, vision-capable, and broadly available across providers, it's a dependable general-purpose workhorse for chat, summarization and light reasoning. Quality is well ahead of the smallest open models while pricing stays modest, which makes it a common…

Google's fastest, lowest-cost hosted Gemini tier. A very large context window and very low latency make it well suited to high-volume hosted workloads where you'd rather not manage providers yourself. Multimodal input, reliable uptime, and first-party serving only. A strong choice if you want a managed endpoint and don't need the routing…

The larger sibling of gpt-oss-20b: a 117B-parameter open-weight MoE (5.1B active) under Apache 2.0. Noticeably stronger reasoning, coding and instruction-following while staying inexpensive to serve thanks to the sparse architecture. Hosted by a dozen providers for competitive live pricing and failover. A common open default when the 20B…

Mistral's compact open model, refreshed for better tool use and structured output. Apache-2.0 and widely hosted, it's even-tempered and predictable for everyday tasks like chat, extraction and routing. Not the strongest reasoner in its price band, but reliable and easy to deploy, with competitive live pricing across several providers.

The smaller, faster member of Meta's Llama 4 family. An open-weight MoE with a big context window and native vision. Cheaper and snappier than Maverick, it's a good open option for high-volume chat and document tasks where you want long context without a premium price. Widely hosted, so live prices stay competitive.

A fast, low-cost distill of DeepSeek V4 with native reasoning and a large context window. It punches well above its price on math and coding, and the reasoning mode can be toggled per request. Keep in mind that reasoning tokens count toward output cost, so heavy chain-of-thought usage raises the effective price. Hosted by several…

Meta's open-weight Llama 4 Maverick, a 400B-parameter sparse MoE (17B active) with a very large context window and native image understanding. Strong general reasoning and multilingual coverage, hosted by a wide range of providers so live prices stay competitive. A solid open option when you want frontier-ish breadth without…

Alibaba's native vision-language dense model with a linear-attention mechanism for fast responses. Performance approaches the much larger Qwen3.5-122B-A10B on many benchmarks, and it handles long inputs comfortably with a 262K context window. Multilingual and well-rounded for both text and image tasks, with broad provider coverage that…