llms.txt
Compare with other sites' llms.txt files →# Inference.net
> Inference infrastructure for AI-native teams: serverless and dedicated model hosting, LLM tracing and observability, evaluations, and fine-tuning, behind an OpenAI-compatible API.
Each link below is the Markdown copy of a page on https://inference.net; append `.md` to any page URL to get one (the home page is `/index.md`).
## Docs
- [Inference.net documentation](https://docs.inference.net/llms.txt): API reference, SDK guides, and platform docs, with their own llms.txt index.
## Product
- [Home](https://inference.net/index.md): Inference infrastructure for AI-native teams: serve, monitor, evaluate, and train production AI models.
- [Deploy](https://inference.net/products/deploy.md): Dedicated, autoscaling model hosting with low latency and 99.99% uptime.
- [Train](https://inference.net/products/train.md): Turn production data into specialized models with lower cost and latency.
- [Observe](https://inference.net/products/observe.md): Latency, cost, quality, and behavior for every model call across providers.
- [Trace](https://inference.net/products/trace.md): OpenTelemetry tracing for agents: LLM calls, tool calls, and framework steps.
- [Evaluate](https://inference.net/products/evaluate.md): Evaluate models on real production traffic before you ship.
- [Pricing](https://inference.net/pricing.md): Plans, included usage, and limits for the Pay as you go, Growth, and Enterprise plans.
- [Model catalog](https://inference.net/models.md): Every public model with context size, capabilities, and per-token pricing.
- [Privileged access](https://inference.net/privileged-access.md): Models open only to approved teams, and how to request access to them.
## Resources
- [Blog](https://inference.net/blog.md): News, research, and technical posts from the Inference.net team.
- [Articles](https://inference.net/content.md): Articles on LLM training, deployment, observability, and evaluation.
- [Guides](https://inference.net/guides.md): Step-by-step tutorials for building on Inference.net.
- [Case studies](https://inference.net/case-studies.md): How teams train, deploy, and optimize LLMs on Inference.net.
- [Inference.net vs Langfuse](https://inference.net/compare/langfuse.md): Feature comparison, pricing math, and migration guide from Langfuse.
## Company
- [Company](https://inference.net/company.md): The team and mission behind Inference.net.
- [Brand guidelines](https://inference.net/brand.md): Logo downloads, colors, and typography.
- [Enterprise](https://inference.net/enterprise.md): Dedicated support, custom SLAs, and secure infrastructure for large teams.
- [Grants](https://inference.net/grants.md): Up to $10,000 in free compute for open-source AI developers and researchers.
- [Meet with us](https://inference.net/meet-with-us.md): Book a call with the Inference.net research team.
## Legal
- [Privacy policy](https://inference.net/privacy-policy.md): How Inference.net collects, uses, and protects personal data.
- [Terms of service](https://inference.net/terms-of-service.md): Terms that govern use of the Inference.net platform.
- [Partner terms of service](https://inference.net/partner-terms-of-service.md): Terms for the Inference.net Partner Program.
- [Zero data retention policy](https://inference.net/zero-data-retention.md): How Inference.net handles the prompts and outputs of API requests.
## Models
- [Kimi K3](https://inference.net/models/kimi-k3.md): Moonshot AI's flagship open-weight reasoning model: a 2.8T-parameter Mixture-of-Experts that activates 104B parameters per token, with a 1M-token context window, native vision, and configurable reasoning effort. Frontier-grade agentic coding served on an OpenAI-compatible API at a fraction of frontier pricing.
- [MiniMax M3.1 Flash](https://inference.net/models/minimax-m3.1-flash.md): MiniMax's multimodal reasoning model for agentic coding, computer use, and cybersecurity: a 428B-parameter Mixture-of-Experts that activates 23B parameters per token, with a 1M-token context window, image input, and five reasoning-effort levels. Near-frontier agentic performance at Flash-tier pricing.
- [Nemotron 3 Super](https://inference.net/models/nemotron-3-super.md): Nemotron 3 Super is a high-throughput, open-weight 120B hybrid mixture-of-experts model by NVIDIA with 12B active parameters, optimized for complex agentic AI workflows. Featuring a 1-million-token context window, hybrid Mamba-transformer architecture, and multi-token prediction, it is designed for scalable deployment on workstations, data centers, and cloud environments.
- [Schematron V2 Small](https://inference.net/models/schematron-v2-small.md): A specialized extraction model that turns messy HTML into clean, schema-conforming JSON. Schematron V2 Small delivers the family's highest extraction quality — approaching frontier-model accuracy on complex schemas and long pages at a small fraction of the cost.
- [Schematron V2 Turbo](https://inference.net/models/schematron-v2-turbo.md): The high-throughput tier of the Schematron family, purpose-built to turn messy HTML into schema-conforming JSON. Optimized for speed and cost at scale — pass a JSON Schema and a page, get back clean, typed data over a 128K context.
- [DeepSeek V4.1 Flash 1017 Cyber](https://inference.net/models/deepseek-v4.1-flash-1017-cyber.md): A refusal-removed variant of the October 2026 DeepSeek V4.1 Flash checkpoint. Same 552B-parameter multimodal Mixture-of-Experts with image input, tool calling, and a 1M-token context. Available only to teams approved for privileged model access.
- [Claude Fable 5](https://inference.net/models/claude-fable-5.md): Anthropic's frontier model, built for the hardest reasoning and long-horizon agentic work. A 1M-token context window and adaptive thinking let it sustain complex, multi-step tasks — deep research, large-scale refactors, autonomous coding sessions — that shorter-context models can't hold together.
- [Claude Haiku 4.5](https://inference.net/models/claude-haiku-4-5.md): Anthropic's small, fast Claude — near-Sonnet quality at a fraction of the price. The right default for high-volume production work: classification, extraction, summarization, and low-latency chat where cost per call is the binding constraint.
- [Claude Opus 4.6](https://inference.net/models/claude-opus-4-6.md): The first Opus-class Claude with a 1M-token context window. Opus 4.6 plans carefully, sustains long agentic sessions, and works reliably inside large codebases, with adaptive thinking for trading depth against speed on each request.
- [Claude Opus 4.7](https://inference.net/models/claude-opus-4-7.md): An Opus-class Claude tuned for software engineering and multi-step agentic work, with a 1M-token context window and extended thinking. Superseded by Opus 4.8 for new projects but still served for teams pinned to this version in production.
- [Claude Opus 4.8](https://inference.net/models/claude-opus-4-8.md): Anthropic's workhorse flagship for agentic coding and knowledge work. Opus 4.8 catches subtle code flaws that earlier Opus models let pass, holds a 1M-token context, and exposes adaptive thinking for balancing intelligence, latency, and cost.
- [Claude Opus 5](https://inference.net/models/claude-opus-5.md): A step-change upgrade to the Opus tier that approaches Claude Fable 5's frontier intelligence at a significantly lower price. Thinking is available across a 1M-token context with configurable effort for scaling test-time compute up or down per request.
- [Claude Opus 5.5](https://inference.net/models/claude-opus-5-5.md): Anthropic's successor to Claude Opus 5 at a lower price: a 1M-token context with image input, tool calling, structured outputs, and always-on adaptive reasoning with effort from low up to max. Opus-tier coding and agentic quality at $4 / $20 per million tokens.
- [Claude Sonnet 4.6](https://inference.net/models/claude-sonnet-4-6.md): Anthropic's best balance of speed and intelligence. Sonnet 4.6 covers most production workloads — chat, document analysis, moderate-complexity coding — with a 1M-token context window and adaptive thinking at well below Opus pricing.
- [DeepSeek V3.2](https://inference.net/models/deepseek-v3.2.md): DeepSeek's cost-efficient V3.2 reasoning model: hybrid thinking that trades depth against speed on a 163K-token context, with tool calling, structured outputs, and JSON mode. Open-weight quality at commodity pricing.
- [DeepSeek V4 Flash](https://inference.net/models/deepseek-v4-flash.md): The efficient tier of DeepSeek's open-weight V4 family: a sparse Mixture-of-Experts that activates only a small fraction of its parameters per token, with a 1M-token context and selectable reasoning effort. Frontier-adjacent quality at a fraction of frontier cost.
- [DeepSeek V4 Flash (0731)](https://inference.net/models/deepseek-v4-flash-0731.md): The July 2026 refresh of DeepSeek V4 Flash, served as a distinct pinned version. Same efficient Mixture-of-Experts design with a 1M-token context, plus low, high, and max reasoning modes for dialing test-time compute per request.
- [DeepSeek V4 Pro](https://inference.net/models/deepseek-v4-pro.md): DeepSeek's flagship open-weight reasoning model: a large sparse Mixture-of-Experts with a 1M-token context window. At high reasoning effort it ranks among the strongest open models on math and coding, at pricing far below closed frontier peers.
- [DeepSeek V4 Pro (0813)](https://inference.net/models/deepseek-v4-pro-0813.md): The August 2026 refresh of DeepSeek V4 Pro, served as a distinct pinned version. Same flagship sparse Mixture-of-Experts with a 1M-token context, 131K max output, and low, high, and max reasoning modes for dialing test-time compute per request.
- [DeepSeek V4.1 Flash](https://inference.net/models/deepseek-v4.1-flash.md): DeepSeek's efficient V4.1 Flash tier: a natively multimodal 552B-parameter sparse Mixture-of-Experts tuned for coding, cybersecurity, and agentic workloads, with image input, tool calling, a 1M-token context, and a continuously controllable reasoning-effort dial. Near-frontier agentic quality at a small fraction of frontier pricing.
- [DeepSeek V4.1 Flash 1017](https://inference.net/models/deepseek-v4.1-flash-1017.md): The October 2026 checkpoint of DeepSeek V4.1 Flash: the same 552B-parameter multimodal Mixture-of-Experts, with stronger reasoning, coding, agentic, and visual results and shorter outputs that finish tasks faster at lower cost. Image input, tool calling, and a controllable reasoning-effort dial.
- [Gemini 2.5 Flash](https://inference.net/models/gemini-2.5-flash.md): Google's proven price-performance workhorse. Gemini 2.5 Flash takes text, image, audio, and video across a 1M-token context with controllable thinking — a dependable default for summarization, chat, and high-volume multimodal pipelines.
- [Gemini 2.5 Flash Lite](https://inference.net/models/gemini-2.5-flash-lite.md): The fastest, cheapest model in the Gemini 2.5 line. Flash Lite is built for latency-sensitive, high-throughput work — classification, translation, routing — while keeping the full 1M-token multimodal context.
- [Gemini 2.5 Pro](https://inference.net/models/gemini-2.5-pro.md): Google's Gemini 2.5 reasoning flagship. Strong on math, code, and long-document analysis across a 1M-token context, with full multimodal input — text, images, audio, and video — and controllable thinking budgets.
- [Gemini 3 Flash Preview](https://inference.net/models/gemini-3-flash-preview.md): The preview build that introduced the Gemini 3 Flash line: near-frontier intelligence at Flash speed and pricing, over a 1M-token multimodal context. Served for developers pinned to the preview; new work should target Gemini 3.5 or 3.6 Flash.
- [Gemini 3.1 Flash Lite](https://inference.net/models/gemini-3.1-flash-lite.md): The low-cost, high-throughput tier of Google's Gemini 3.1 generation. Built for volume — subagent fleets, bulk document extraction, structured parsing — across a 1M-token multimodal context at the lowest price in its family.
- [Gemini 3.1 Pro Preview](https://inference.net/models/gemini-3.1-pro-preview.md): Google's Gemini 3 frontier reasoning model, refined for factual grounding and precise tool use. Optimized for software engineering and multi-step agentic workflows over a 1M-token context with text, image, audio, video, and PDF input.
- [Gemini 3.5 Flash](https://inference.net/models/gemini-3.5-flash.md): Google's most intelligent Flash model: frontier-level performance on agentic execution, coding, and long-horizon tasks at Flash speed and pricing. A 1M-token multimodal context and dynamic thinking make it a strong default for fast agentic loops.
- [Gemini 3.5 Flash-Lite](https://inference.net/models/gemini-3.5-flash-lite.md): The fastest model in the Gemini 3.5 series, designed for high-throughput execution — subagents, bulk extraction, structured parsing — with a 1M-token multimodal context at the lowest price point in the family.
- [Gemini 3.6 Flash](https://inference.net/models/gemini-3.6-flash.md): Google's workhorse Flash model, tuned for agentic coding and token efficiency: it completes multi-step workflows with fewer output tokens and fewer tool calls than 3.5 Flash, at a lower price, across the same 1M-token multimodal context.
- [Gemini 3.7 Flash](https://inference.net/models/gemini-3.7-flash.md): Google's latest and most capable Flash model, built for complex coding, agentic workflows, and reliable multi-step execution. Same 1M-token multimodal context as 3.6 Flash, sharper reasoning and tool use, at Flash pricing.
- [Gemma 3 27B](https://inference.net/models/gemma-3-27b-it.md): Google's open-weight 27B instruction-tuned model. Multimodal (text and image input) with a 128K context and broad multilingual coverage — a capable, permissively licensed workhorse for chat, summarization, and vision-language tasks.
- [Gemma 3 4B](https://inference.net/models/gemma-3-4b-it.md): Google's compact open-weight Gemma: the 4B instruction-tuned tier, multimodal (text and image input) with a 128K context. Built for on-device-adjacent workloads — routing, classification, drafting — where unit cost and latency dominate.
- [GLM-5](https://inference.net/models/glm-5.md): Z.ai's open-weight GLM-5, MIT-licensed, with a 200K-token context and a thinking toggle for reasoning-heavy prompts. An economical default for chat, extraction, and agentic tool use when the newer 5.2/5.3 tiers are more than a task needs.
- [GLM-5.2](https://inference.net/models/glm-5.2.md): Z.ai's flagship open-weight model for long-horizon reasoning and agentic tasks, MIT-licensed with a 1M-token context. Sparse attention keeps long-context inference efficient, and it posts near-frontier results on reasoning, coding, and agentic benchmarks.
- [GLM-5.2 Fast](https://inference.net/models/glm-5.2-fast.md): GLM-5.2 on Inference.net's low-latency serving tier, with the same open weights, 1M-token context, reasoning, structured outputs, and tool support. Use it for interactive agents that need faster generation.
- [GLM-5.3](https://inference.net/models/glm-5.3.md): Z.ai's GLM-5.3: an open-weight reasoning model for long-horizon agentic and coding tasks, with a 1M-token context and 131K max output. Served on Fireworks' serverless tier with prompt caching, structured outputs, and tool use.
- [GLM-5.3 Flash](https://inference.net/models/glm-5.3-flash.md): Z.ai's GLM-5.3 Flash: a fast open-weight reasoning model with image input, a 1M-token context, and 131K max output. Reasoning effort is selectable from none through max, with prompt caching for high-volume agentic and coding workloads.
- [GPT-4.1](https://inference.net/models/gpt-4.1.md): OpenAI's dependable non-reasoning workhorse, tuned for instruction following and coding, with a 1M-token context window. No thinking tokens means predictable latency and cost — a solid default for production text and vision workloads.
- [GPT-4.1 Mini](https://inference.net/models/gpt-4.1-mini.md): A mid-size GPT-4.1 that matches or beats GPT-4o on many tasks at a fraction of the cost and latency, with the same 1M-token context. A strong balance point for chat, extraction, and tool calling at scale.
- [GPT-4.1 Nano](https://inference.net/models/gpt-4.1-nano.md): OpenAI's fastest, cheapest GPT-4.1 tier. Built for classification, autocomplete, and high-volume extraction where latency and unit cost dominate, while keeping the 1M-token context and image input.
- [GPT-4o](https://inference.net/models/gpt-4o.md): OpenAI's battle-tested multimodal generalist. GPT-4o handles text and image input with reliable function calling and structured outputs over a 128K context — a proven default for chat and vision workloads.
- [GPT-4o Mini](https://inference.net/models/gpt-4o-mini.md): The cost-efficient GPT-4o tier that made high-volume LLM pipelines affordable. Text and vision input, function calling, and structured outputs over a 128K context — ideal for chatbots, extraction, and bulk processing.
- [GPT-5 Mini](https://inference.net/models/gpt-5-mini.md): A compact GPT-5 reasoning model that balances capability against cost. Configurable reasoning effort from minimal to high across a 400K context makes it a strong default for production agents that need reasoning without flagship pricing.
- [GPT-5 Nano](https://inference.net/models/gpt-5-nano.md): The smallest, fastest GPT-5 tier. Reasoning effort dials down to minimal for near-instant responses, with a 400K context and image input — built for high-throughput classification, extraction, and routing.
- [GPT-5.4](https://inference.net/models/gpt-5.4.md): OpenAI's early-2026 frontier reasoning model with a 1M-token context and reasoning effort from none to xhigh. A capable all-rounder for coding and knowledge work, now priced below its GPT-5.5 and 5.6 successors.
- [GPT-5.5](https://inference.net/models/gpt-5.5.md): OpenAI's agentic specialist and the backbone of Codex-style coding workflows. GPT-5.5 excels at terminal work and multi-step tool use, holds a 1M-token context, and completes long agentic loops with fewer interruptions than its predecessors.
- [GPT-5.6 Luna](https://inference.net/models/gpt-5.6-luna.md): The fastest, most affordable member of the GPT-5.6 family. Luna handles long-running professional workflows at a fraction of flagship cost, with a 1M-token context and the full none-to-max reasoning ladder.
- [GPT-5.6 Sol](https://inference.net/models/gpt-5.6-sol.md): OpenAI's flagship. Sol leads the GPT-5.6 family across coding, knowledge work, and science, completing tasks with fewer output tokens than earlier frontier models — with a 1M-token context and reasoning effort from none up to max.
- [GPT-5.6 Terra](https://inference.net/models/gpt-5.6-terra.md): The balanced everyday tier of the GPT-5.6 family: performance competitive with GPT-5.5 at a significantly lower price. A 1M-token context and none-to-max reasoning efforts make it a strong default for production agents.
- [GPT-6 Astra](https://inference.net/models/gpt-6-astra.md): OpenAI's GPT-6 flagship for the hardest agentic coding and knowledge-work tasks: a 922K-token context with image input, tool calling, structured outputs, and reasoning effort from none up to xhigh. Built for deep multi-step work where quality dominates cost.
- [GPT-6 Luna](https://inference.net/models/gpt-6-luna.md): OpenAI's cost-efficient GPT-6 model for high-volume workloads: a 922K-token context with image input, tool calling, structured outputs, and reasoning effort from none up to xhigh. Delivers GPT-6-class intelligence at a fraction of the cost, making it the default for cost-sensitive production agents.
- [GPT-6 Sol](https://inference.net/models/gpt-6-sol.md): OpenAI's balanced GPT-6 model for coding and agentic workflows: a 922K-token context with image input, tool calling, structured outputs, and reasoning effort from none up to xhigh. Delivers near-flagship intelligence at half the cost of GPT-6 Astra, making it the default for high-volume production agents.
- [GPT