⚡Stack Radar

AI & Automation · AI Models & Inference

Inference APIs

9 paid against 9 free you can run yourself.

Paid against free › AI & Automation › AI Models & Inference › Inference APIs · AI and models on Repo Radar

Inference APIs let you call AI models over an API and pay per token or per second, without running GPUs yourself. The paid providers differ in catalog and price (DeepInfra and Novita are among the cheapest for open models, Together and Fireworks add fine-tuning and dedicated capacity, Cerebras sells speed), in billing (per token, per second of hardware on Replicate and Baseten) and in reach (OpenRouter and Hugging Face route one key to many providers). On the open-source side you can run your own OpenAI-compatible API with vLLM, SGLang, LocalAI or Xinference, run models locally with Ollama or llama.cpp, or self-host a gateway such as LiteLLM. Marks come from vendor pricing pages and project READMEs as of 18 October 2026; nothing was tested hands-on.

Paid 9 reviewed

Prices and free tiers as checked, dated on each review.

DDeepInfra

Production-ready inference APIs for open-source AI models with pay-as-you-go pricing.

free tierTTogether AI

Full-stack AI cloud for building, fine-tuning, and deploying open-source models.

free tierFFireworks AI

Fast, affordable AI inference and training platform for open models.

free tierOOpenRouter

Unified API for 500+ AI models with routing, fallbacks, and pay-as-you-go pricing.

free tierRReplicate

Cloud API to run and fine-tune open-source AI models.

free tierBBaseten

Deploy and scale AI models in production.

free tierCCerebras

Fast AI inference and training on wafer-scale hardware.

free tierNNovita AI

Unified API for 200+ AI models, GPU instances, and agent runtimes.

free tierHHugging Face Inference Providers

Unified API for serverless AI inference across multiple providers.

$9free tier

Free 9 direct · 12 related

Free to use. You run it on your own server.

Direct alternatives

They do this job.

LLggml-org/llama.cpp

A C and C++ library and command-line tool for running language models locally.

★ 132kFreeVLvllm-project/vllm

A high-throughput and memory-efficient inference and serving engine for LLMs

★ 94.1kFreeLIBerriAI/litellm

A gateway that puts a hundred or more model providers behind one interface, with spend tracking and limits.

★ 61kFreeLOmudler/LocalAI

Serves text, vision, speech, image and video models from one self-hosted service that imitates the paid interfaces.

★ 47.8kFreeSGsgl-project/sglang

SGLang is a high-performance serving framework for large language models and multimodal models.

★ 37.3kFreeFRtashfeenahmed/freellmapi

One local address in front of the free tiers of many AI providers.

★ 36.6kFreeINxorbitsai/inference

Swap GPT for any LLM by changing a single line of code. Xinference lets you run open-source, speech, and multimodal models on cloud, on-prem, or your laptop — all through one unified, production-ready inference API.

★ 9.6kFreeDYai-dynamo/dynamo

A Datacenter Scale Distributed Inference Serving Framework

★ 8.3kFreeMAmagnitudedev/magnitude

It measures the machine, tunes itself to that hardware, then runs open AI models locally.

★ 6.7kFree

Related alternatives

A second choice: they cover part of this job or are a building block for it.

OLollama/ollama

It downloads and runs AI language models on the computer in front of you.

★ 183kFreeTRhuggingface/transformers

🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.

★ 168kFreeLLhiyouga/LlamaFactory

Unified Efficient Fine-Tuning of 100+ LLMs & VLMs (ACL 2024)

★ 75.6kFreeLLmozilla-ai/llamafile

A whole language model and the program that runs it, in one downloadable file.

★ 26.4kFreeGPopenai/gpt-oss

gpt-oss-120b and gpt-oss-20b are two open-weight language models by OpenAI

★ 20.5kFreeAPcaronc/apprise

One way to send a message to any of a hundred different services.

★ 17.6kFreeLILibreTranslate/LibreTranslate

Self-hosted machine-translation API — removes the per-character meter from running a bilingual site.

★ 17.1kFreeOUdottxt-ai/outlines

Forces an AI model's answer into a shape you define, while the answer is being produced.

★ 15.9kFreeKIMoonshotAI/Kimi-K3

Open Frontier Intelligence

★ 8.9kFreeHEHelicone/helicone

🧊 Open source LLM observability platform. One line of code to monitor, evaluate, and experiment. YC W23 🍓

★ 6.2kFreeGEgoogle-deepmind/gemma

Gemma open-weight LLM library, from Google DeepMind

★ 5.8kFreeQWQwenLM/Qwen3.8

Qwen3.8 is the large language model series developed by Qwen team, Alibaba Group.

★ 4.3kFree

Feature by feature

Every paid tool and free repository we compared, marked against the same features.

✓has itPartlysome plans or extra work—confirmed not available?not confirmed✓the vendor's claim onlypricingClick a mark for its note and source

Paid · 9Free · 10
Featurevllm-project/vllmsgl-project/sglangmudler/LocalAIollama/ollamaggml-org/llama.cppBerriAI/litellmxorbitsai/inferencemagnitudedev/magnitudeai-dynamo/dynamotashfeenahmed/freellmapi
Starts atPay as you go per token; no subscriptionPay as you go per token; no subscriptionPay as you go per token; no subscription from prepaid creditsProvider list prices plus 5.5% fee on credits (Standard)Pay as you go; billed per output, token or second of hardware$0 per month, pay as you goSelf-serve pay as you go (Developer tier)Pay as you go per token; no subscription$9 per month (Pro)Apache-2.0Apache-2.0MITMITMITMITApache-2.0Apache-2.0Apache-2.0MIT
Models
Open-weight models per tokenServerless access to open-weight models such as DeepSeek, Qwen or Kimi, billed per token.
Integration
OpenAI-compatible APIWorks with code written for the OpenAI API by changing the base URL.
Models
Many providers in one APIOne API key reaches models from many providers, including closed models.
Pricing
Prompt caching discountCached input tokens cost less than fresh ones.
Batch API discountCheaper rates for jobs that can wait.
Capacity
Dedicated endpoints or GPUsReserved deployments or GPU instances instead of shared serverless capacity.
Models
Fine-tuningTrain your own version of a model and serve it.
Image, video or audio modelsGeneration or recognition models beyond text.
Embedding modelsModels that turn text into vectors for search.
Pricing
Free credits or free tierSome usage free to start, as credits or free models.
Integration
Routing and your own keysRoute across providers or bring your own provider keys.
Teams
Enterprise SSO or RBACSingle sign-on or role-based access for organisations.
Budgets and spend controlsLimits and budgets on spending.
Deployment
Self-hostingRun the inference service on your own machines.
Runs on consumer hardwareRuns on laptops, CPUs or Apple silicon, not only data-centre GPUs.
Distributed multi-GPU servingServes one model across several GPUs or machines.

Excluded candidates

Looked at, and left out of the comparison, with the reason.

Paid apps

  • Groqgroq.com/pricing now redirects to the Groq homepage, which describes its inference cloud but publishes no per-token price list, so its prices cannot be compared with the selected providers.Source

Repositories

  • huggingface/text-generation-inferenceHugging Face's Text Generation Inference server is archived, with its last push on 21 March 2026, so it is not a supported choice today.Source
  • containers/ramalamaRamaLama serves AI models locally inside containers. It covers the same local workflow as Ollama and LocalAI with a smaller community, so it ranked below them.Source
  • kvcache-ai/ktransformersKTransformers is a framework for experimenting with mixed CPU-GPU inference and fine-tuning optimisations. It is a research toolkit rather than a ready inference API server.Source
  • bentoml/BentoMLBentoML is a framework for packaging and serving any AI app or model as an API. It is general model-serving tooling, broader than an inference API for language models.Source
  • Portkey-AI/gatewayPortkey Gateway routes to many LLM providers with guardrails, like LiteLLM, but its last push was on 25 May 2026, so LiteLLM was preferred.Source
  • NVIDIA/TensorRT-LLMTensorRT-LLM is an NVIDIA library for optimising models for its GPUs. It is an engine component used underneath servers such as Dynamo rather than an inference API itself.Source
  • bentoml/OpenLLMOpenLLM runs open models as OpenAI-compatible APIs. It overlaps vLLM, Xinference and LocalAI, which are more active, so it ranked below them.Source
  • jundot/omlxoMLX is an inference server for Apple silicon with continuous batching and SSD caching. It is Mac-only, narrower than llama.cpp and Ollama, which also run on Macs.Source
  • xtekky/gpt4freegpt4free collects access to powerful models through third-party websites. It relies on services that did not agree to that use, so it is not a dependable inference API.Source
  • mnfst/llm-gatewayAn open-source LLM gateway with about 7,600 stars. It overlaps LiteLLM, which supports more providers and is more widely used, so it ranked below it.Source
  • ray-project/rayRay is a general distributed compute engine; Ray Serve can host models, but it is infrastructure for building such services rather than an inference API.Source

Gaps

What the research could not settle from the vendors' own pages.

What vendors don't publish

  • Pricing pages list per-token prices but not the rate limits that come with them, so two providers with the same price can deliver very different throughput.DeepInfra, Together AI, Novita AI · checked 2026-10-18 · source 1 · source 2 · source 3
  • Replicate bills many models by run time on the chosen hardware rather than per token, so its costs cannot be compared directly with per-token providers without running a workload.Replicate · checked 2026-10-18 · source 1
  • Hugging Face publishes the monthly inference credit on its pricing page but each partner provider's token prices only in the documentation, so the real cost per call is not on one page.Hugging Face Inference Providers · checked 2026-10-18 · source 1 · source 2

Paid against free, in your inbox

One job at a time: the paid tools worth paying for, the free ones you can run yourself, and who should pick which.

No tracking pixels. One click to leave. We use your address to send Stack Radar and nothing else. Privacy.