Open-source frontier intelligence.
Zyphra and frontier open-source models you can select, adapt and consume in the cloud or an environment you control.
Recent models on Zyphra Inference
Featuring Zyphra’s Optimization Engine (ZOE).
GLM-5.3-Flash (FP4)
- 1M-token context
- OpenAI-compatible API
- FP4 weights & FP8 KV cache
- 213
- tokens/s/userPer-user decode with 32 sessions
- 62%
- lower output pricingvs Together AI, Nebius and Modal
USD per 1M tokens
| Provider | Input | Cached input | Output |
|---|---|---|---|
| $0.11 | $0.02 | $0.19 | |
| $0.15 | $0.03 | $0.50 | |
![]() | $0.15 | $0.03 | $0.50 |
![]() | $0.15 | $0.03 | $0.50 |
- 9.5×
- higher per-user output speed¹
- 48×
- faster TTFT¹
¹ GLM-5.3-Flash, 8x MI300X, SemiAnalysis AgentX agentic coding benchmark, 96 concurrent sessions. Stock SGLang v0.5.22 with MTP speculative decoding. Internal benchmark November 13, 2026.
GLM-5.3-Flash
- 1M-token context
- OpenAI-compatible API
- FP8 weights & FP8 KV cache
- 156
- tokens/s/userPer-user decode with 32 sessions
- 50%
- lower output pricingvs Z.ai, Fireworks and Together.ai
USD per 1M tokens
| Provider | Input | Cached input | Output |
|---|---|---|---|
| $0.15 | $0.03 | $0.25 | |
| ZAI | $0.15 | $0.03 | $0.50 |
| $0.15 | $0.03 | $0.50 | |
| $0.15 | $0.03 | $0.50 |
- 3.1×
- higher per-user output speed¹
- 10.7×
- faster TTFT¹
¹ GLM-5.3-Flash, 8x MI300X, SemiAnalysis AgentX agentic coding benchmark, 64 concurrent sessions. Stock SGLang v0.5.22 with MTP speculative decoding. Internal benchmark November 4, 2026.
Optimized frontier open-source endpoints
Call models through the Zyphra Inference API, have Zyphra run managed inference for you, or license ZOE to serve open models on your own compute.
Zyphra Inference API
Use an API key to call supported models from your application, through an OpenAI-compatible interface.
Managed inference
Put open models behind your products without operating the model-serving infrastructure. Zyphra runs model serving, with ZOE optimizing it.
License ZOE
Run open models on your own accelerators, with model- and hardware-specific optimization from ZOE.
Open-source frontier models optimized with Zyphra’s Optimization Engine
Zyphra and frontier open-source models served by Zyphra with an OpenAI-compatible API. Optimized with Zyphra’s Optimization Engine to deliver maximum intelligence and performance.
Zyphra models
- Parameters
- 886M · 100M active
- Context
- 32K
- Price
- $10.00 / 1M UTF-8 bytes
- Parameters
- 380M
- Price
- Free
- Parameters
- 8B · 900M active
- Context
- 6144
- Price
- $20.00 / 1M UTF-8 bytes
- Parameters
- 8B · 760M active
- Context
- 128K
- Price
- $0.00 in · $0.00 out
Open-source third-party models
ZAI · Language- Parameters
- 320B · 18B active
- Context
- 1M
- Price
- $0.11 in · $0.19 out
ZAI · Language- Parameters
- 320B · 18B active
- Context
- 1M
- Price
- $0.15 in · $0.25 out
Simple, transparent pricing.
Language models
Price per 1M tokens, in USDAudio and thought models
Price per unit, in USDTell us what you’re building.
Bring your workload, model requirements and operating environment. We’ll discuss the managed inference or ZOE license that fits.


