Frontier open models for coding agents
Call GLM-5.3, Kimi K3, and DeepSeek-V4-Flash through an OpenAI-compatible API. Fast enough for apply loops. Priced by the token, not the seat.
Works with your coding agent
01 / PRODUCT INTERFACE
Deploy AI agents with one API call.
Stop harness-engineering. Bring your system prompt, skills, tools, and connectors, then get infrastructure, orchestration, memory, self learning, and more out of the box. Then connect over SSE or use directly in Slack.
Reusable versioned configuration
Read docsDurable threads and runs
Read docsTokens and tool calls over SSE
Read docsUsage and events for your logs
Read docsLearns every user
Read docsScheduled background runs
Read docsMarkdown playbooks on demand
Read docsWeb, files, and shell built in
Read docsYour tools, open protocol
Read docsDMs and channel mentions
Read docsA persistent workspace
Read docsCredentials, never in the sandbox
Read docsSend /goal and keep going
Read docsIsolated per agent and subject
Read docsImproves between sessions
Read docsGate tool calls by policy
Read docs02 / Models
Three open-weight models. One API.
GLM-5.3, Kimi K3, and DeepSeek-V4-Flash for the primary agent loop. Long context, token pricing, no per-seat fee.
GLM-5.3
Sparse MoE for agentic coding, reasoning, and tool use. The default loop model when you want long context without a closed-weight lock-in.
prism-glm53 · Agent-loop throughput
View GLM-5.3Kimi K3
2.8T MoE with roughly 50B active parameters. Use it when the agent needs vision, deep reasoning, and a long-lived conversation cache.
prism-kimik3 · Reasoning-first
View Kimi K3DeepSeek-V4-Flash
Fast MoE with compressed attention. The cheap, low-latency open-weight option for inner-loop coding, reasoning, and tool use.
prism-dsv4flash · Up to 150 tok/s
View DeepSeek-V4-Flash03 / Drop-in API
Point your OpenAI client at Prism.
Same chat completions shape. Change the base URL and model id. Auth is PRISM_API_KEY.
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api.herm.run/v1",
apiKey: process.env.PRISM_API_KEY,
});
const completion = await client.chat.completions.create({
model: "prism-glm53",
messages,
});04 / Inference stack
Build your agent, not your inference stack.
Serving open models yourself means vLLM, a router, KV cache sizing, and latency work before the first apply loop. Prism is the model API. Point your existing client at it and keep shipping the agent.
One key covers GLM-5.3, Kimi K3, and DeepSeek-V4-Flash. Swap the model id when the loop needs more context or more speed.
Get an API key05 / Questions
What is Prism?
Prism is an OpenAI-compatible inference API for coding agents. It serves open-weight frontier models so you can run the primary agent loop without standing up your own GPU stack.
Which models can I call?
The public lineup is GLM-5.3, Kimi K3, and DeepSeek-V4-Flash. Model ids are prism-glm53, prism-kimik3, and prism-dsv4flash.
Is it OpenAI-compatible?
Yes. Point any OpenAI SDK at https://api.herm.run/v1 and call chat completions. Keep your existing client. Change the base URL, API key, and model id.
How do I get an API key?
Book a demo at cal.com/team/prismai/demo. After the call, set PRISM_API_KEY and send it with your requests.
How is inference priced?
Usage is billed per token. There is no per-seat fee. See Pricing for credit plans, or talk to us for dedicated capacity.
Which coding agents work with Prism?
Any client that can set an OpenAI-compatible base URL: Cursor, Codex, Claude Code, Openclaw, Hermes, and custom loops. Copy the setup prompt on the homepage if you want an agent to wire it for you.
Can I get a private deployment?
Yes. Book a call for dedicated capacity in our cloud or yours.
06 / Get a key