Introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens.

DeepSeek-V4.1-Flash

GLM-5.3 is a large-scale reasoning model from Z.ai, built for complex software engineering and long-horizon agent tasks. It supports text input and output with a 1M-token context window, and improves on GLM-5.2 in coding and in the balance between performance and token efficiency.

GLM-5.3

GLM-5.3-Flash is a native multimodal model from Z.ai. It is suited for efficient coding and long-horizon agent tasks. Its hybrid sparse and linear attention architecture maintains accurate long-context behavior while reducing compute overhead.

GLM-5.3-Flash

Kimi K3 is Moonshot AI's 2.8T-parameter open-weight multimodal reasoning model. Built for complex coding, knowledge work, and long-horizon agentic workflows, it excels at navigating large repositories, calling tools, debugging, and iterating on images, logs, tests, and runtime feedback. Its architecture relies on KDA and Attention Residuals for computational efficiency.

Kimi-K3

Qwen3.8 2.4T A95B is an open-weight sparse mixture-of-experts model from Qwen and the open-weight variant of Qwen3.8 Max, with 95 billion active parameters out of 2.4 trillion total. It is suited for coding, research, complex reasoning, and agentic workflows.

Qwen3.8-2.4T-A95B

DeepSeek-V4-Pro-0813 is the official release of DeepSeek-V4-Pro, superseding the preview version, with greatly enhanced agentic capabilities and performance improvements that are especially pronounced in production environments. It is built on the DeepSeek-V4-Pro (Preview) model structure, with a DSpark speculative decoding module attached.

DeepSeek-V4-Pro-0813

DeepSeek-V4-Flash-0731 is the official release of DeepSeek-V4-Flash, superseding the preview version, with substantially enhanced agentic capabilities. DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available.

DeepSeek-V4-Flash-0731

GLM-5.2 is Z-AI's latest flagship model for long-horizon tasks. It marks a substantial leap in long-horizon task capability over its predecessor GLM-5.1 and, for the first time, delivers that capability on a **solid 1M-token context**.

GLM-5.2

Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With substantial improvements on real-world long-horizon coding tasks, it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6.

Kimi-K2.7-Code

Nemotron 3 Ultra is built for, frontier reasoning, orchestration, coding agents, deep research, and complex enterprise workflows. It delivers up to 5x faster inference and up to 30% lower cost for agentic workloads while supporting up to 1M token context.

NVIDIA-Nemotron-3-Ultra-550B-A55B

DeepSeek V4 Flash is an efficiency-focused MoE model with 284B total parameters (13B active) and a 1M-token context window. It's tuned for fast inference and high-throughput use cases while still holding up on reasoning and coding tasks.

DeepSeek-V4-Flash

DeepSeek V4 Pro is an MoE model with 1.6T total parameters (49B active) and a 1M-token context window. It's built for advanced reasoning, coding, and long-running agent tasks, and performs well on knowledge, math, and software engineering benchmarks.

DeepSeek-V4-Pro

Kimi K2.6 is an open-source, native multimodal agentic model that advances practical capabilities in long-horizon coding, coding-driven design, proactive autonomous execution, and swarm-based task orchestration.

Kimi-K2.6

MiMo-V2.5-Pro is an open-source Mixture-of-Experts (MoE) language model with 1.02T total parameters and 42B active parameters. It utilizes the hybrid attention architecture and 3-layers Multi-Token Prediction (MTP) introduced in [MiMo-V2-Flash](https://github.com/XiaomiMiMo/MiMo-V2-Flash).

MiMo-V2.5-Pro

Qwen3.6-35B-A3B is Alibaba's latest flagship Mixture-of-Experts model, with 35B total parameters and only 3B activated per token (256 experts, 8 routed + 1 shared). Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience.

Qwen3.6-35B-A3B

GLM-5.1 is Z-AI's next-generation flagship model for agentic engineering, with significantly stronger coding capabilities than its predecessor. It achieves state-of-the-art performance on SWE-Bench Pro and leads GLM-5 by a wide margin on NL2Repo (repo generation) and Terminal-Bench 2.0 (real-world terminal tasks).

GLM-5.1

Qwen3.5-397B-A17B is Alibaba's most capable Qwen3.5 model, a Mixture-of-Experts architecture with 397B total parameters and 17B activated per token. It features a 262K token context window (extensible to 1M with YaRN), thinking/reasoning mode, tool calling with MCP integration, and support for 201 languages. Sets state-of-the-art results on reasoning, coding, math, and multimodal benchmarks.

Qwen3.5-397B-A17B

Efficient, MoE variant of Gemma 4. Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input and generating text output.

gemma-4-26B-A4B-it

Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input and generating text output.

gemma-4-31B-it

NVIDIA Nemotron 3 Super is a hybrid Mixture-of-Experts (MoE) model engineered for highest compute efficiency and accuracy in multi-agent applications and specialized agentic systems. It is optimized to run many collaborating agents per application on a single GPU, delivering high accuracy for reasoning, tool use, and instruction following.

NVIDIA-Nemotron-3-Super-120B-A12B

  Qwen3-TTS is an advanced text-to-speech model by Alibaba's Qwen team, delivering stable, expressive, and low-latency speech generation across 10 languages.                                                                                                                                                                                                                                                                                                                                           Key capabilities:                                                                                                                                                                                                                                  - 9 preset voices — Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee — covering diverse genders, ages, and accents                                                                                                              - Voice cloning — clone any voice from a short (~3s) audio sample via the voice_id parameter   - Instruction control — adjust tone, emotion, and speaking style with natural language (e.g. "speak slowly and calmly", "excited tone")   - 10 languages — English, Chinese, Japanese, Korean, German, French, Russian, Spanish, Italian, Portuguese   - Streaming support — real-time PCM streaming with ~97ms first-byte latency   - Multiple output formats — WAV, MP3, FLAC, PCM    Built on a 1.7B parameter architecture using discrete multi-codebook language modeling for end-to-end speech synthesis without cascading errors. Uses a custom 12Hz acoustic tokenizer that preserves paralinguistic information and   environmental audio details.

Qwen3-TTS

● Qwen3-TTS-VoiceDesign is a voice design variant of Qwen3-TTS by Alibaba's Qwen team. Instead of selecting from preset voices, you describe the voice you want in natural language — and the model generates speech in that voice.                                                                                                                                                                                                                                                                     Key capabilities:                                                                                                                                                                                                                                  - Natural language voice control — describe any voice with free text (e.g. "a deep male voice with a calm, authoritative presence", "a young cheerful female with a warm and friendly tone")   - 10 languages — English, Chinese, Japanese, Korean, German, French, Russian, Spanish, Italian, Portuguese                                                                                                                                         - Streaming support — real-time PCM streaming   - Multiple output formats — WAV, MP3, FLAC, PCM    Built on the same 1.7B parameter architecture as Qwen3-TTS, using discrete multi-codebook language modeling and a custom 12Hz acoustic tokenizer for high-quality end-to-end speech synthesis.

Qwen3-TTS-VoiceDesign

The latest flagship model in the Qwen family. State-of-the-art results across a comprehensive suite of benchmarks — including knowledge, reasoning, coding, instruction following, human preference alignment, agent tasks, and multilingual understanding.

Qwen3-Max

The latest flagship reasoning model in the Qwen3 family. Further enhanced by multiple innovations like adaptive tool-use and advanced test-time scaling techniques

Qwen3-Max-Thinking

DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism that reduces training and inference cost while preserving quality in long-context scenarios. A scalable reinforcement learning post-training framework further improves reasoning, with reported performance in the GPT-5 class, and the model has demonstrated gold-medal results on the 2025 IMO and IOI. V3.2 also uses a large-scale agentic task synthesis pipeline to better integrate reasoning into tool-use settings, boosting compliance and generalization in interactive environments.

DeepSeek-V3.2

The fastest model of the Flux 2 family. Frontier visual intelligence — state-of-the-art image generation and editing from Black Forest Labs

FLUX-2-klein-4b

The best quality-to-latency ratio, production apps model of the Flux 2 family. Frontier visual intelligence — state-of-the-art image generation and editing from Black Forest Labs

FLUX-2-klein-9b

claude

Claude

deepseek

DeepSeek

flux

Flux

gemini

Gemini

kimi

Kimi

llama

Llama

mistral

Mistral

nemotron

Nemotron

qwen

Qwen


# Understanding LoRA inference

### Concepts

- **Base model**: The original model that is used as a starting point.
- **LoRA adapter model**: A small model that is used to adapt the base model for a specific task.
- **LoRA Rank**: The rank of the matrix that is used to adapt the model.

### What you need to inference with LoRA adapter model

1. Supported base model
2. LoRA adapter model hosted on HuggingFace
3. HuggingFace token if the LoRA adapter model is private
4. DeepInfra account

### How to inference with LoRA adapter in DeepInfra

1. Go to the [dashboard](https://deepinfra.com/dash)
2. Click on the 'New Deployment' button
3. Click on the 'LoRA Model' tab
4. Fill the form:
    - **LoRA model name**: model name used to reference the deployment
    - **Hugging Face Model Name**: Hugging Face model name
    - **Hugging Face Token**: (optional) Hugging Face token if the LoRA adapter model is private
5. Click on the 'Upload' button

Note: The list of supported base models is listed on the same page. If you need a base model that is not listed, please contact us at [feedback@deepinfra.com](mailto:feedback@deepinfra.com)

### Rate limits on LoRA adapter model

Rate limit will apply on combined traffic of all LoRA adapter models with the same base model. For example, if you have 2 LoRA adapter models with the same base model, and have rate limit of 200. Those 2 LoRA adapter models combined will have rate limit of 200.

### Pricing on LoRA adapter model

Pricing is 50% higher than base model.

### How is LoRA adapter model speed compared to base model speed?

LoRA adapter model speed is lower than base model, because there is additional compute and memory overhead to apply the LoRA adapter. From our benchmarks, the LoRA adapter model speed is about 50-60% slower than base model.

### How to make LoRA adapter model faster?

You could merge the LoRA adapter with the base model to reduce the overhead. And use custom deployment, the speed will be close to the base model.
