MorphMorph
Morph Models

Fast general models for agent loops

Run the primary agent loop on fast, OpenAI-compatible coding models served on Morph's custom kernels. One API for chat, code generation, and reasoning.
Get API KeyRead the docsDedicated endpoints
Same coding job on two stacks: baseline serving grinds line by line while Kimi K3 on Morph snaps every edit in and settles at 100 tok/s
Morph Models

Frontier coding models, served on custom kernels

Output speed

Codegen-specific optimizations and custom GPU kernels. Up to 150 tok/s on DeepSeek V4 Flash, measured on private deployments with custom speculators. Public endpoints run on shared capacity.

tok/s on private deployments150DSV4 Flash150Kimi K3 Fast100Kimi K3100Qwen 3.6 27B80GLM-5.2 744B90MiniMax M3

The lineup

Open-weight frontier models with long context, served and billed per token. No per-seat fees.

01
// Available general models
02
morph-kimik3           // 2.8T MoE, 1M context
03
morph-kimik3-fast      // same weights, latency-tuned
04
morph-glm52-744b       // 744B MoE, 1M context
05
morph-minimax3-428b    // 428B MoE, 256k context
06
morph-dsv4flash        // 1M context, fast
07
morph-qwen38-27b       // dense, low latency

Built for agent loops, not chat

Long tool calls expire your cache before the next turn arrives. Tag turns with run_id and the scheduler keeps it hot.

Docs
01
// One field. Every turn of a run shares it.
02
await client.chat.completions.create({
03
  model: "morph-kimik3",
04
  messages,
05
  run_id: "run-8f2c1a",          // sticky worker + priority resume
06
  // parent_run_id: "run-root",  // subagent fan-out lineage
07
});
08
// run_final: true releases the run when the loop ends (not billed)

Standby: half price when nobody is waiting

service_tier: "standby" pays 50% of GLM-5.2's rate. Runs on spare capacity, sheds a fast 429 when busy — batch jobs, evals, and nightly runs skip interactive pricing.

Docs
01
// GLM-5.2 standby — $0.55/M in, $0.11/M cached, $2.05/M out
02
await client.chat.completions.create({
03
  model: "morph-glm52-744b",
04
  messages,
05
  service_tier: "standby",  // 50% off; 429 + Retry-After when busy
06
});
07
// Rejections generate nothing and bill nothing. Retry with
08
// backoff, or resend as "default" if it can't wait.

Built for production agent workloads

Inference optimized for coding agents

Every agent will write code. We bet the stack on it.

So we tune every layer for that one workload: custom GPU kernels, speculative decoding shaped around code, and serving built for the agent loop instead of general chat. Not general infrastructure with code bolted on.

Private deployments

The fastest endpoints are private deployments

The headline speeds on this page come from dedicated deployments, not the shared public endpoints. At scale we build the deployment around your traffic: speculators trained on it, caching tuned to it, and pricing well under public rates.

100B+

tokens per day served across private deployments

Custom speculators

trained on your traffic for higher acceptance rates

Custom caching

prefix and KV caching tuned to your workload

Volume pricing

large discounts over public per-token rates

View self-serve commitmentsTalk to us about a custom deployment