Hosted Whisper and Parakeet transcription.
22 integrated models · Official documentation ↗
DeepSeek-V4.1-Flash (September 2026), the first model on DeepSeek's V4.1 architecture: it uses a 552B backbone activating 8B parameters in prefill and 16B in decode, and supports 1M tokens under MIT.
GLM-5.3-Flash (August 2026), a newly trained Z.ai GLM-5 model: a 320B MoE with 18B active, hybrid sparse and linear attention and MIT weights, which Z.ai says beats GLM-5.2 at about a tenth of the price.
Z.ai's GLM-5.3 (August 2026), GLM-5.2's base with heavier post-training for coding and agentic work and a 1M-token context, served by Alibaba Cloud, Mistral and Together.
The August 2026 general-availability release of DeepSeek-V4-Pro, a 1.6T-parameter MoE with 49B active and a 1M-token context, served by DeepInfra and Together with low, high and max reasoning effort.
Muse Glimmer (August 2026, Apache 2.0), Meta Superintelligence Labs' 30B dense open model distilled from Muse Spark for agentic tool use, served with a 128K context by DeepInfra and Together.
The July 2026 release of DeepSeek-V4-Flash (284B MoE, 13B active, 1M-token context), served by DeepInfra and Together; DeepSeek's V4.1-Flash has since succeeded it.
Thinking Machines Lab's Inkling (July 2026, Apache 2.0) on DeepInfra and Together: a 975B-parameter MoE with 41B active that reasons with variable effort, aimed at agentic and tool-use work.
Kimi K3 (July 2026), Moonshot AI's flagship open-weight model, served by DeepInfra and Together: a 2.8T-parameter MoE with 104B active and a 1M-token context for large-repository coding and agent work.
Prism ML's Ternary Bonsai 27B (July 2026): Qwen3.6-27B compressed to ternary {-1, 0, +1} weights (about 1.7 bits each, ~7 GB) while keeping roughly 95% of FP16 benchmark scores and a 262K context. Bonsai 2 27B, built on Qwen3.8-27B, followed in September.
Z.ai's June 2026 open-weight (MIT) model for long-horizon coding, the first GLM to deliver a usable 1M-token context, with 128K output and selectable reasoning effort. Superseded as flagship by GLM-5.3, a post-trained update of the same base.
NVIDIA Nemotron 3.5 ASR (June 2026), a 600M streaming speech recognizer covering 40 language-locales from one checkpoint via language-ID prompting, with automatic language detection, punctuation and selectable latency chunks; weights use OpenMDW-1.1.
MiniMax-M3 (June 2026), MiniMax's flagship open-weight MoE (about 428B total, 23B active) with sparse attention and adaptive thinking, served at a 512K context by DeepInfra and Together.
Qwen3.7-Plus (June 2026), Alibaba's proprietary mid-tier model with a 1M-token context; as of October 2026 it is still Alibaba's recommended Plus model, alongside Qwen3.8-Max and Qwen3.8-Flash.
Qwen3.6-Plus (April 2026), with large gains over 3.5 in agentic and front-end coding; 1M context, equal to the 2026-04-02 snapshot. Superseded by qwen3.7-plus.
Qwen3.5-9B (March 2026), the largest of Alibaba's small dense Qwen3.5 models, on DeepInfra and Together: thinking mode and a 262K context for low-cost and local-scale deployments, under Apache 2.0.
Together's Nemotron 3 ASR Streaming 0.6B is NVIDIA's English Nemotron ASR Streaming model: a 600M cache-aware FastConformer-RNNT with punctuation and chunk sizes from 80 ms to 1.12 s. For other languages NVIDIA points to Nemotron 3.5 ASR.
NVIDIA Parakeet TDT 0.6B v3 (August 2025), a 600M-parameter CC BY 4.0 transcription model for 25 European languages with automatic language detection, punctuation and word timestamps; it does not translate. NVIDIA's newer streaming option is Nemotron 3.5 ASR.
OpenAI's larger open-weight reasoning model (August 2025, Apache 2.0), served here by DeepInfra, Together and Groq: a 117B MoE with 5.1B active, configurable reasoning effort and a 128K context.
Canopy Labs' Orpheus 3B (March 2025), an Apache 2.0 English TTS model on a Llama 3B backbone with eight voices (tara, leah, jess, leo, dan, mia, zac, zoe) and inline emotive tags like <laugh> and <sigh>; streaming latency is about 200 ms.
Kokoro-82M v1.0 (January 2025), hexgrad's Apache 2.0 text-to-speech model derived from StyleTTS 2; at 82M parameters it is small and fast, with 54 voices across 8 languages.
