Models / Open weights

Fast general models for agent loops

Run the primary agent loop on open-weight coding models through one OpenAI-compatible Prism API. Chat, code generation, and reasoning. No closed-weight lock-in.

Built for agent loops, not chatOpen source only1M context

Get an API key

01 / The lineup

Open-weight frontier models with long context, served and billed per token. No per-seat fees. The dropdown models are the only models on this page.

Qwen

Qwen is the dense 27B in the lineup: low latency, open weights, and enough coding quality for tight apply loops without a giant MoE.

Model ID

prism-qwen

Architecture

27B dense

Context

256K

27B dense checkpoint. Fast to load and cheap to serve.
Low-latency decode for inner-loop coding and tool use.
Open weights on the same OpenAI-compatible Prism endpoint.

One field. Every turn of a run shares it.

const completion = await client.chat.completions.create({
  model: "prism-qwen",
  messages,
  prompt_cache_key: "run-8f2c1a",
});

Available general models

// Available open-source models
prism-glm53           // 744B MoE, 1M context
prism-kimik3          // 2.8T MoE, 1M context
prism-dsv4flash       // fast MoE, 1M context
prism-qwen            // 27B dense, low latency

Point any OpenAI SDK at the Prism API. One key covers Qwen and the rest of the open-source lineup.

02 / Built for agent loops

Output speed

Codegen-specific serving for the inner loop. DeepSeek-V4-Flash is the fast lane; GLM-5.3 and Kimi K3 hold the longer reasoning traces.

Keep a conversation on its cache

Untagged turns scatter across workers and re-prefill the whole prompt. One cache key pins every turn to the worker that already holds it.

Open weights only

This page is the open-source lineup. Closed models stay off the models nav so the catalog you see is the catalog you can self-host.

03 / Production workloads

Built for production agent workloads

Every agent will write code. Prism tunes serving for that workload: long tool traces, sticky caches, and open-weight models you can run on Prism or take with you.

Low latency

Speculative decoding and batching shaped around code generation, not general chat.

High throughput

Batched serving across a GPU fleet so a customer agent can stay in a tight loop.

Private deployments

Dedicated capacity with caching tuned to your traffic, in our cloud or yours.

04 / Deploy

The cheapest and fastest inference on the market.

Get an API key Book a demo