
Open-weight multimodal agent model tuned with large-scale RL and a 1M-token window, served on Deepinfra.
Deepinfra gives developers simple, quick APIs for running open models at production scale without overspending.
DeepSeek-V4.1-Flash-1017
NVIDIA-Nemotron-3.5-Super-VL
On-Demand DGX B300 GPUs
Hy4-preview
MiMo-V2.6-Flash
GLM-5.3
Kimi-K3.1-Pro
Funding news · May 4, 2026
The round was led jointly by 500 Global and Georges Harik, with NVIDIA, Felicis, Samsung Next, Supermicro, Crescent Cove and several other backers joining. Deepinfra is putting it toward capacity and new hardware.
$107M
Series B raised
25x
more tokens served since Series A
Usage-based billing with no lock-in and no surprise line items. Two founders in a garage or a global enterprise — Deepinfra grows with you, through clean APIs and engineers who actually answer.
Tune for latency, throughput, cost or sheer volume. Deepinfra builds the deployment around whichever matters most to you, with more than 100 hosted models ready to call.
Prompts, completions and end-user data are never stored. Deepinfra is certified for ISO 27001 and SOC 2, and treats security reviews as routine work rather than a yearly scramble.
Deepinfra operates its own GPU fleet inside secured US facilities, tuned end to end for serving models — so latency stays low and uptime stays high.
Models
Speech and image models sit in the full Deepinfra catalogue. Open it here

Open-weight multimodal agent model tuned with large-scale RL and a 1M-token window, served on Deepinfra.

Incremental refresh of V4.1-Flash with sharper reasoning, coding and agent scores, plus shorter answers.

120B mixture-of-experts vision model (12B active) that reads images, documents, screens and video, then acts.

Early look at Tencent's 770B-parameter mixture-of-experts flagship, callable through the Deepinfra API.

Efficiency-focused sparse MoE from Xiaomi's V2.6 line, built for quick, lightweight everyday calls.

Top model of Xiaomi's V2.6 family: a 1.02T-parameter sparse MoE aimed at hard reasoning work.

Multimodal MoE on a 552B backbone, balancing speed and depth for agents and chat on Deepinfra.

Z.ai's large reasoning model for complex software engineering and long-horizon agent tasks.

Natively multimodal sibling of GLM-5.3, sized for quick coding help and long agent runs.

Moonshot's 2.8T-parameter open-weight multimodal reasoner for demanding coding jobs, hosted on Deepinfra.

Official V4-Pro release replacing the preview build, with broad gains in reasoning quality.

Official V4-Flash release that supersedes the preview, keeping latency low for high-volume traffic.
Speed, scale, stability and spend across the Deepinfra fleet, refreshed as traffic flows.
40.42M
Tokens / second
345ms
Time to first token
3654
Requests / second
24.35
exaFLOPS online
DeepCluster
Hardware sourced, racked and run by Deepinfra in a Tier 3 facility. You own the capacity; we keep it healthy under a 99.982% uptime SLA.
Plan your clusterNVIDIA B300 · 5-year term
99.982%
uptime, written into the SLA