DeepInfra

The VL extension of Ling-3.0-flash (September 2026), served here for text input only; 124B MoE with about 5.5B active and a context of up to 1M tokens, thinking on by default.

Mid-size Granite 4.2 model (dense 8B, August 2026) from IBM, with built-in chain-of-thought that can be switched to low-effort or off, tool calling for agent workflows and a 128K native context; Apache 2.0, aimed at cost-conscious enterprise use.

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.

Ant Group's Ling-3.0-flash (August 2026): a hybrid-linear MoE (Kimi Delta Attention plus MLA) with 124B total and 5.1B active parameters and hybrid reasoning, tuned for token-efficient agentic work. Succeeded by the much larger Ling-3.1-flash in October 2026.

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.

Audio8 TTS Preview 0.6B (July 2026) from Edge0: an open Apache 2.0 text-to-speech model with zero-shot voice cloning, a 44.1 kHz codec and a dual autoregressive design inspired by Fish Audio S2; preview release covering 11 languages.

Compact Nemotron 3 embedding model (about 1.1B parameters, July 2026) pruned and distilled from a Ministral 3 3B embedder, giving 2048-dimension multilingual vectors from 32K-token inputs at much lower cost than the 8B; full BF16 precision.

NVFP4-quantized build of Nemotron-3-Embed-1B for Blackwell GPUs (July 2026), retaining near-BF16 retrieval quality with the same 2048-dimension multilingual vectors and 32K-token inputs at lower memory and compute.

NVIDIA's larger Nemotron 3 text embedding model (July 2026), built on Ministral 3 8B with bidirectional attention: 4096-dimension vectors that can be truncated, 32K-token inputs, and strong multilingual and cross-lingual retrieval for RAG.

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.

Smaller sibling of Inkling (July 2026): a 276B MoE with 12B active parameters from Thinking Machines Lab, keeping variable thinking effort at much lower serving cost; Apache 2.0.

FastVideo's 3-step quantization-aware distillation of Wan2.1-T2V-1.3B with FP8 layers (June 2026), producing 5-second 480p clips at 16 fps in a few seconds; trading visual detail for speed against larger Wan models.

NVIDIA Nemotron 3.5 ASR (June 2026): a 600M-parameter cache-aware FastConformer-RNNT model that transcribes 40 language-locales from one checkpoint, built for low-latency streaming captions and voice agents as well as batch transcription.

Largest model of NVIDIA's open Nemotron 3 family (June 2026): a 550B LatentMoE with 55B active, hybrid Mamba-2/attention layers, multi-token prediction and NVFP4 pre-training, aimed at complex agents, coding and long-context reasoning; NVIDIA cites up to 1M-token context.

NVIDIA's Nemotron 3.5 Content Safety guard model (June 2026), a LoRA fine-tune of Gemma 3 4B that classifies content as safe or unsafe against built-in or custom policies, with an optional reasoning trace; multilingual, served here for text input only.

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.

April 2026 preview of DeepSeek-V4-Flash, the efficient tier of the V4 family: 284B-parameter MoE with 13B active, a 1M-token context and selectable reasoning effort. Replaced by the official V4-Flash-0731 and then V4.1-Flash.

April 2026 preview of DeepSeek-V4-Pro, the large tier of the V4 family: a 1.6T-parameter MoE with 49B active, 1M-token context and hybrid compressed attention, aimed at reasoning, coding and long-running agents. Replaced by the GA release V4-Pro-0813.