Over the past few months, we have observed increasingly clear trends toward scaling both total parameters and context lengths in the pursuit of more powerful and agentic artificial intelligence (AI). We are excited to share our latest advancements in addressing these demands, centered on improving scaling efficiency through innovative model architecture. We call this next-generation foundation models Qwen3-Next.

Qwen3-Next-80B-A3B-Instruct

Qwen3-Next-80B-A3B-Thinking

Kimi K2 0905 is the September update of Kimi K2 0711. It is a large-scale Mixture-of-Experts (MoE) language model developed by Moonshot AI, featuring 1 trillion total parameters with 32 billion active per forward pass. It supports long-context inference up to 256k tokens, extended from the previous 128k.  This update improves agentic coding with higher accuracy and better generalization across scaffolds, and enhances frontend coding with more aesthetic and functional outputs for web, 3D, and related tasks. Kimi K2 is optimized for agentic capabilities, including advanced tool use, reasoning, and code synthesis. It excels across coding (LiveCodeBench, SWE-bench), reasoning (ZebraLogic, GPQA), and tool-use (Tau2, AceBench) benchmarks. The model is trained with a novel stack incorporating the MuonClip optimizer for stable large-scale MoE training.

Kimi-K2-Instruct-0905

DeepSeek-V3.1 is post-trained on the top of DeepSeek-V3.1-Base, which is built upon the original V3 base checkpoint through a two-phase long context extension approach, following the methodology outlined in the original DeepSeek-V3 report. We have expanded our dataset by collecting additional long documents and substantially extending both training phases. The 32K extension phase has been increased 10-fold to 630B tokens, while the 128K extension phase has been extended by 3.3x to 209B tokens. Additionally, DeepSeek-V3.1 is trained using the UE8M0 FP8 scale data format to ensure compatibility with microscaling data formats.

DeepSeek-V3.1

gpt-oss-120b is an open-weight, 117B-parameter Mixture-of-Experts (MoE) language model from OpenAI designed for high-reasoning, agentic, and general-purpose production use cases. The model supports configurable reasoning depth, full chain-of-thought access, and native tool use, including function calling, browsing, and structured output generation.

gpt-oss-120b

gpt-oss-20b is an open-weight 21B parameter model released by OpenAI under the Apache 2.0 license. It uses a Mixture-of-Experts (MoE) architecture with 3.6B active parameters per forward pass, optimized for lower-latency inference. The model is trained in OpenAI’s Harmony response format and supports reasoning level configuration, fine-tuning, and agentic capabilities including function calling, tool use, and structured outputs.

gpt-oss-20b

Qwen3-Coder-480B-A35B-Instruct is the Qwen3's most agentic code model, featuring Significant Performance on Agentic Coding, Agentic Browser-Use and other foundational coding tasks, achieving results comparable to Claude Sonnet.

Qwen3-Coder-480B-A35B-Instruct-Turbo

The GLM-4.5 series models are foundation models designed for intelligent agents. GLM-4.5 has 355 billion total parameters with 32 billion active parameters, while GLM-4.5-Air adopts a more compact design with 106 billion total parameters and 12 billion active parameters. GLM-4.5 models unify reasoning, coding, and intelligent agent capabilities to meet the complex demands of intelligent agent applications.

GLM-4.5

olmOCR is a specialized AI tool that converts PDF documents into clean, structured text while preserving important formatting and layout information. What makes olmOCR particularly valuable for developers is its ability to handle challenging PDFs that traditional OCR tools struggle with—including complex layouts, poor-quality scans, handwritten text, and documents with mixed content types. Built on a fine-tuned 7B vision-language model, olmOCR provides enterprise-grade PDF processing at a fraction of the cost of proprietary solutions.

olmOCR-7B-0725-FP8

Qwen3-235B-A22B-Thinking-2507 is the Qwen3's new model with scaling the thinking capability of Qwen3-235B-A22B, improving both the quality and depth of reasoning. 

Qwen3-235B-A22B-Thinking-2507

Qwen3-Coder-480B-A35B-Instruct

GLM-4.5-Air

Voxtral Small is an enhancement of Mistral Small 3, incorporating state-of-the-art audio input capabilities while retaining best-in-class text performance. It excels at speech transcription, translation and audio understanding.

Voxtral-Small-24B-2507

Voxtral Mini is an enhancement of Ministral 3B, incorporating state-of-the-art audio input capabilities while retaining best-in-class text performance. It excels at speech transcription, translation and audio understanding.

Voxtral-Mini-3B-2507

The DeepSeek R1 0528 turbo model is a state of the art reasoning model that can generate very quick responses

DeepSeek-R1-0528-Turbo

Qwen3-235B-A22B-Instruct-2507 is the updated version of the Qwen3-235B-A22B non-thinking mode, featuring Significant improvements in general capabilities, including instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage.  

Qwen3-235B-A22B-Instruct-2507

Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support

Qwen3-30B-A3B

Qwen3-32B

Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support. 

Qwen3-14B

The Llama 4 collection of models are natively multimodal AI models that enable text and multimodal experiences. These models leverage a mixture-of-experts architecture to offer industry-leading performance in text and image understanding. Llama 4 Maverick, a 17 billion parameter model with 128 experts

Llama-4-Maverick-17B-128E-Instruct-FP8

The Llama 4 collection of models are natively multimodal AI models that enable text and multimodal experiences. These models leverage a mixture-of-experts architecture to offer industry-leading performance in text and image understanding. Llama 4 Scout, a 17 billion parameter model with 16 experts

Llama-4-Scout-17B-16E-Instruct

The DeepSeek R1 model has undergone a minor version upgrade, with the current version being DeepSeek-R1-0528.

DeepSeek-R1-0528

DeepSeek-V3-0324, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token, an improved iteration over DeepSeek-V3.

DeepSeek-V3-0324

Mistral-Small-3.2-24B-Instruct is a drop-in upgrade over the 3.1 release, with markedly better instruction following, roughly half the infinite-generation errors, and a more robust function-calling interface—while otherwise matching or slightly improving on all previous text and vision benchmarks.

Mistral-Small-3.2-24B-Instruct-2506

Llama Guard 4 is a natively multimodal safety classifier with 12 billion parameters trained jointly on text and multiple images. Llama Guard 4 is a dense architecture pruned from the Llama 4 Scout pre-trained model and fine-tuned for content safety classification. Similar to previous versions, it can be used to classify content in both LLM inputs (prompt classification) and in LLM responses (response classification). It itself acts as an LLM: it generates text in its output that indicates whether a given prompt or response is safe or unsafe, and if unsafe, it also lists the content categories violated.

Llama-Guard-4-12B

QwQ is the reasoning model of the Qwen series. Compared with conventional instruction-tuned models, QwQ, which is capable of thinking and reasoning, can achieve significantly enhanced performance in downstream tasks, especially hard problems. QwQ-32B is the medium-sized reasoning model, which is capable of achieving competitive performance against state-of-the-art reasoning models, e.g., DeepSeek-R1, o1-mini.

QwQ-32B

Anthropic’s most powerful model yet and the state-of-the-art coding model. It delivers sustained performance on long-running tasks that require focused effort and thousands of steps, significantly expanding what AI agents can solve. Claude Opus 4 is ideal for powering frontier agent products and features.

claude-4-opus

Anthropic's mid-size model with superior intelligence for high-volume uses in coding, in-depth research, agents, & more.

claude-4-sonnet

Gemini 2.5 Flash is Google's latest thinking model, designed to tackle increasingly complex problems. It's capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy.  Gemini 2.5 Flash: best for balancing reasoning and speed.

gemini-2.5-flash

Gemini 2.5 Pro is Google's the most advanced thinking model, designed to tackle increasingly complex problems. Gemini 2.5 Pro leads common benchmarks by meaningful margins and showcases strong reasoning and code capabilities.  Gemini 2.5 models are thinking models, capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy.  The Gemini 2.5 Pro model is now available on DeepInfra.

gemini-2.5-pro

Gemma 3 introduces multimodality, supporting vision-language input and text outputs. It handles context windows up to 128k tokens, understands over 140 languages, and offers improved math, reasoning, and chat capabilities, including structured outputs and function calling. Gemma 3 27B is Google's latest open source model, successor to Gemma 2

gemma-3-27b-it

Gemma 3 introduces multimodality, supporting vision-language input and text outputs. It handles context windows up to 128k tokens, understands over 140 languages, and offers improved math, reasoning, and chat capabilities, including structured outputs and function calling. Gemma 3-12B is Google's latest open source model, successor to Gemma 2

gemma-3-12b-it

gemma-3-4b-it

Kokoro is an open-weight TTS model with 82 million parameters. Despite its lightweight architecture, it delivers comparable quality to larger models while being significantly faster and more cost-efficient. With Apache-licensed weights, Kokoro can be deployed anywhere from production environments to personal projects.

Kokoro-82M

Orpheus TTS is a state-of-the-art, Llama-based Speech-LLM designed for high-quality, empathetic text-to-speech generation. This model has been finetuned to deliver human-level speech synthesis, achieving exceptional clarity, expressiveness, and real-time streaming performances.

orpheus-3b-0.1-ft

CSM (Conversational Speech Model) is a speech generation model from Sesame that generates RVQ audio codes from text and audio inputs. The model architecture employs a Llama backbone and a smaller audio decoder that produces Mimi audio codes.

csm-1b

DeepSeek-R1-Distill-Llama-70B is a highly efficient language model that leverages knowledge distillation to achieve state-of-the-art performance. This model distills the reasoning patterns of larger models into a smaller, more agile architecture, resulting in exceptional results on benchmarks like AIME 2024, MATH-500, and LiveCodeBench. With 70 billion parameters, DeepSeek-R1-Distill-Llama-70B offers a unique balance of accuracy and efficiency, making it an ideal choice for a wide range of natural language processing tasks. 

DeepSeek-R1-Distill-Llama-70B

DeepSeek-V3, a strong Mixture-of-Experts (MoE) language model with 671B total parameters with 37B activated for each token. To achieve efficient inference and cost-effective training, DeepSeek-V3 adopts Multi-head Latent Attention (MLA) and DeepSeekMoE architectures, which were thoroughly validated in DeepSeek-V2. 

DeepSeek-V3

Llama 3.3-70B Turbo is a highly optimized version of the Llama 3.3-70B model, utilizing FP8 quantization to deliver significantly faster inference speeds with a minor trade-off in accuracy. The model is designed to be helpful, safe, and flexible, with a focus on responsible deployment and mitigating potential risks such as bias, toxicity, and misinformation. It achieves state-of-the-art performance on various benchmarks, including conversational tasks, language translation, and text generation.

Llama-3.3-70B-Instruct-Turbo

Llama 3.3-70B is a multilingual LLM trained on a massive dataset of 15 trillion tokens, fine-tuned for instruction-following and conversational dialogue. The model is designed to be helpful, safe, and flexible, with a focus on responsible deployment and mitigating potential risks such as bias, toxicity, and misinformation. It achieves state-of-the-art performance on various benchmarks, including conversational tasks, language translation, and text generation.

Llama-3.3-70B-Instruct

Phi-4 is a model built upon a blend of synthetic datasets, data from filtered public domain websites, and acquired academic books and Q&A datasets. The goal of this approach was to ensure that small capable models were trained with data focused on high quality and advanced reasoning.

phi-4

Whisper is a state-of-the-art model for automatic speech recognition (ASR) and speech translation, proposed in the paper "Robust Speech Recognition via Large-Scale Weak Supervision" by Alec Radford  et al. from OpenAI. Trained on >5M hours of labeled data, Whisper demonstrates a strong ability to generalise to many datasets and domains in a zero-shot setting. Whisper large-v3-turbo is a finetuned version of a pruned Whisper large-v3. In other words, it's the exact same model, except that the number of decoding layers have reduced from 32 to 4. As a result, the model is way faster, at the expense of a minor quality degradation.

whisper-large-v3-turbo

BGE embedding is a general Embedding Model. It is pre-trained using retromae and trained on large-scale pair data using contrastive learning. Note that the goal of pre-training is to reconstruct the text, and the pre-trained model cannot be used for similarity calculation directly, it needs to be fine-tuned

DeepInfra supports the OpenAI embeddings API.
The following creates an embedding vector representing the input text

```bash
curl "https://api.deepinfra.com/v1/openai/embeddings" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPINFRA_TOKEN" \
  -d '{
    "input": "The food was delicious and the waiter...",
    "model": "BAAI/bge-base-en-v1.5",
    "encoding_format": "float"
  }'
```

which will return something similar to

```json
{
  "object":"list",
  "data":[
    {
      "object": "embedding",
      "index":0,
      "embedding":[
        -0.010480394586920738,
        -0.0026091758627444506
        ...
        0.031979579478502274,
        0.02021978422999382
      ]
    }
  ],
  "model": "BAAI/bge-base-en-v1.5",
  "usage": {
    "prompt_tokens":12,
    "total_tokens":12
  }
}
```


You can use the official OpenAI python client to create embedding vectors

```python
from openai import OpenAI

# Create an OpenAI client with your deepinfra token and endpoint
openai = OpenAI(
    api_key="$DEEPINFRA_TOKEN",
    base_url="https://api.deepinfra.com/v1/openai",
)

input = "The food was delicious and the waiter...", # or an array ["hello", "world"]

embeddings = openai.embeddings.create(
  model="BAAI/bge-base-en-v1.5",
  input=input,
  encoding_format="float"
)

if isinstance(input, str):
    print(embeddings.data[0].embedding)
else:
    for i in range(len(input)):
        print(embeddings.data[i].embedding)

print(embeddings.usage.prompt_tokens)

# [0.0241, 0.5, 1.0, ..., 0.3123, 0.743564]
# 42
```


You can use the official OpenAI Node.js client to create embedding vectors

```javascript
import OpenAI from "openai";

const openai = new OpenAI({
  baseURL: 'https://api.deepinfra.com/v1/openai',
  apiKey: "$DEEPINFRA_TOKEN",
});

const input = "The quick brown fox jumped over the lazy dog" // or an array ["hello", "world"]

async function main() {
  const embedding = await openai.embeddings.create({
    model: "BAAI/bge-base-en-v1.5",
    input: input,
    encoding_format: "float",
  });

  // check if input is a string or array
  if (typeof input === "string") {
    console.log(embedding.data[0].embedding);
  } else {
    console.log(embedding.data.map((data) => data.embedding));
  }

  console.log(embedding.usage.prompt_tokens);
}

main();

// [0.0241, 0.5, 1.0, ..., 0.3123, 0.743564]
// 42
```


You can use cURL or any other http client to run inferences:

```bash
curl -X POST \
    -d '{"inputs": ["I like chocolate"]}'  \
    -H "Authorization: bearer $DEEPINFRA_TOKEN"  \
    -H 'Content-Type: application/json'  \
    'https://api.deepinfra.com/v1/inference/BAAI/bge-base-en-v1.5'
```

which will give you back something similar to:

```json
{
  "embeddings": [
    [
      0.0,
      0.5,
      1.0
    ],
    [
      1.0,
      0.5,
      0.0
    ]
  ],
  "input_tokens": 42,
  "embedding_jsons": [
    "[0.0, 0.5, 1.0]",
    "[1.0, 0.5, 0.0]"
  ],
  "request_id": null,
  "inference_status": {
    "status": "unknown",
    "runtime_ms": 0,
    "cost": 0.0,
    "tokens_generated": 0,
    "tokens_input": 0
  }
}

```


You can use our command-line tool [deepctl](/docs/advanced/deepctl) to run
inferences:

```bash
deepctl infer \
    -m 'BAAI/bge-base-en-v1.5'  \
    -i 'inputs=["I like chocolate"]'
```

which will give you back something similar to:

```json
{
  "embeddings": [
    [
      0.0,
      0.5,
      1.0
    ],
    [
      1.0,
      0.5,
      0.0
    ]
  ],
  "input_tokens": 42,
  "embedding_jsons": [
    "[0.0, 0.5, 1.0]",
    "[1.0, 0.5, 0.0]"
  ],
  "request_id": null,
  "inference_status": {
    "status": "unknown",
    "runtime_ms": 0,
    "cost": 0.0,
    "tokens_generated": 0,
    "tokens_input": 0
  }
}

```


We recommend using our NodeJS client https://github.com/deepinfra/deepinfra-node.

You can install it with

```bash
npm install deepinfra
```

and then

```javascript
import { Embeddings } from "deepinfra";

const DEEPINFRA_API_KEY = "$DEEPINFRA_TOKEN";
const MODEL = "BAAI/bge-base-en-v1.5";

const main = async () => {
  const client = new Embeddings(MODEL, DEEPINFRA_API_KEY);
  const body = {
    inputs: [
      "What is the capital of France?",
      "What is the capital of Germany?",
      "What is the capital of Italy?",
    ],
  };
  const output = await client.generate(body);
  console.log(output.embeddings[0]);
};

main();
```


The service tier used for processing the request. When set to 'priority', the request will be processed with higher priority.

service_tier

inputs

whether to normalize the computed embeddings

normalize

image

The number of dimensions in the embedding. If not provided, the model's default will be used.If provided bigger than model's default, the embedding will be padded with zeros.

dimensions

The webhook to call when inference is done, by default you will get the output in the response of your inference request

webhook

ServiceTier

Dimensions

Image

Inputs

Normalize

Webhook

EmbeddingsIn

estimated cost billed for the request in USD

Cost

Runtime Ms

Status

Tokens Generated

Tokens Input

InferenceReplyStatus

Embedding Jsons

the computed embeddings for each input sequence or a single embedding if an image was passed