No top-up. No credit card. — No PayPal required — No daily login required — Just completely free models, available instantly via an OpenAI-compatible v1 API. —
Tested twice daily with real requests.
Explore free models directly from the aggregators’ API catalogs: we automatically test every free model with a real request. Only when it responds successfully do we list the model here with its provider and aggregator.
Every model must meet our condition, so you can use it without conditions.
Last API scan · 48 free models from 24 providers via 11 aggregators.
Ampersand v0.1 is a collaborator model from humans& for group conversations and shared goals. It is trained to decide when to contribute, surface relevant context and help people work through disagreements.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
The official humans& model card identifies Ampersand v0.1 and the API identifier ampersand-v0.1. Vercel's official model page uses the same exact humansand/ampersand-v0.1 route ID. The hosted model is not merged with its GLM 5.3 base model.
Gemini 4 Hydrogen is Google's fast, multimodal Gemini 4 model for everyday coding, agentic workflows and multimodal knowledge work. It provides a unified one-million-token context and configurable thinking levels.
Details & thinking
Limits above: Google AI Studio. Available gateway routes are listed below.
Google's official model card and Cloud documentation identify the exact model ID gemini-4-hydrogen and the Google Gemini 4 Hydrogen model. The route is a direct Google listing, so no cross-gateway merge is applied.
Particle-1 is a token-efficient agentic coding model post-trained on DeepSeek V4.1 Flash. It targets software engineering, terminal workflows and data engineering while reducing unnecessary reasoning and tool-call tokens.
Details & thinking
Limits above: Cline. Available gateway routes are listed below.
Cline lists the exact upstream ID particle-ai/particle-1:free. Particle's official launch post identifies Particle-1 as the same model, and the OpenRouter listing uses the same canonical ID. The shared presentation alias keeps the three gateway routes together without changing their provider IDs.
Baidu's text model supports coding, document analysis and scripted 3D modeling using application-supplied tools. This intermediate post-trained checkpoint builds on ERNIE 5.5 Flash Base and activates 14 billion of 280 billion parameters.
Details & thinking
Cline's exact free-route catalog advertises zero input and output token prices. Separately, the supplied successful test at 2026-11-11T12:07:02+00:00 establishes access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas remain unverified.
Limits above: Cline. Available gateway routes are listed below.
Cline's official full catalog explicitly identifies baidu/ernie-5.5-flash-preview:free and maps it to baidu/ERNIE-5.5-Flash-Preview. This matches the supplied reviewed Kilo exact-version entry and Baidu's announcement at https://ernie.baidu.com/release/ernie-5-5-flash-preview/. The shared advertised identity does not establish identical gateway configurations or pinned repository revisions.
Meituan's multimodal coding model combines visual understanding with software engineering across complex codebases. It supports visual coding, creative design and 3D tasks, with multi-step reasoning and tool calls for applications that execute actions and return results.
Details & thinking
Nous explicitly lists LongCat 2.5 among its free models. No successful account-access test or model-specific top-up requirement is supplied; quotas and continued free availability remain unverified.
Limits above: Nous Portal. Available gateway routes are listed below.
Preserved the supplied reviewed identity. Nous's official directory identifies Meituan: LongCat 2.5, corresponding to meituan/longcat-2.5:free. The supplied creator changelog separately announces LongCat-2.5 and LongCat-2.5-Preview; no checkpoint-equivalence mapping establishes that these releases are identical.
Mistral's multimodal model handles instruction following, coding and analysis of images and documents. Its mixture-of-experts architecture activates approximately 52 billion parameters including embeddings and output layers, and includes a draft head for speculative decoding.
Details & thinking
NVIDIA's exact route page advertises a free API endpoint. Separately, the supplied successful test at 2026-11-05T14:18:41+00:00 establishes access without a top-up using the dedicated account that had never held credit or linked a payment method. The schema's possible insufficient-credit error does not establish a mandatory top-up for this successful account. Continued availability and quotas are not established.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA explicitly maps this exact route to Mistral Large 4 and the vendor checkpoint mistralai/Mistral-Large-4-1T-A52B. Mistral's exact-model documentation and repository confirm the identity; no supplied known entry represents this version.
Recursive's Flint is a language model for coding agents that work through repository tasks using application-supplied tools. Its mixture-of-experts architecture activates approximately 2.5 billion of 39.8 billion parameters per token.
Details & thinking
The supplied successful tests for this exact Cline route at 2026-11-05T13:01:51+00:00 and 2026-11-05T18:43:52+00:00 establish access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: Cline. Available gateway routes are listed below.
The supplied Cline recursive/flint-v1.0:free snapshot matches the exact namespaced version in OpenRouter's official catalog and its explicit RecursiveSI/flint repository mapping. This establishes advertised identity without establishing identical gateway configurations or pinned revisions.
MiniMax's multimodal reasoning model is designed for coding, long-document analysis and tasks using application-supplied tools. It combines text, image and video understanding with adjustable thinking depth for more involved work.
Details & thinking
The supplied successful test at 2026-11-05T14:18:24+00:00 establishes access to this exact NVIDIA route without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA's exact minimaxai/minimax-m3.1-flash card explicitly links MiniMaxAI/MiniMax-M3.1-Flash. This establishes the same advertised M3.1 Flash model as the supplied Cline, GMI, Kilo, LarpRouter and OpenRouter entries, without implying identical serving settings or pinned revisions. MiniMax M3 remains distinct.
Reflection's Beam models target coding, reasoning and extended workflows using application-supplied tools. This unversioned Beam alias does not establish the exact checkpoint, so version-specific architecture and weight availability remain unresolved.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
The supplied Kilo catalog records reflection/beam-free. Reflection's official Beam model card establishes the creator and public model family, but no retrieved Kilo source explicitly maps this unversioned alias to Beam-1.0-501B-A23B. Its exact identity remains separate from the versioned Nous route and explicitly mapped OpenCode route.
NVIDIA's vision-language model combines document and video understanding with reasoning, coding and tool use. Its hybrid Mamba-Transformer mixture-of-experts architecture activates 12 billion of 120 billion parameters and incorporates a C-RADIO vision encoder.
Details & thinking
The supplied successful tests for this exact Cline route at 2026-11-05T13:01:25+00:00 and 2026-11-05T18:43:24+00:00 establish access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: Cline. Available gateway routes are listed below.
Preserved the supplied established advertised-model identity shared with Kilo, OpenRouter, NVIDIA and the previously documented OpenCode mapping. The supplied Cline namespaced listing and NVIDIA's exact Super VL card support this identity without establishing identical serving configurations or pinned revisions.
Xiaomi's multimodal model understands text, images, video and audio in one architecture. It combines these inputs with reasoning and tool calls for tasks that span documents, media and code.
Details & thinking
The exact route uses TokenHarbor's free allowance and is not billed; availability, quotas and model access remain account-dependent.
Limits above: TokenHarbor. Available gateway routes are listed below.
https://tokenharbor.ai/models/mimo-v2.5:free explicitly attributes MiMo V2.5 to Xiaomi, matching the supplied reviewed OpenCode MiMo V2.5 entry and https://huggingface.co/XiaomiMiMo/MiMo-V2.5. Reused the existing exact-model duplicate_key without rewriting the known sibling. The access suffix differs by gateway; serving configurations remain separate.
Xiaomi's multimodal model is designed for coding and tool-based tasks across long sessions. It can combine text, images, audio and video while retaining extended code and task histories, rather than handling each input type separately.
Details & thinking
The exact route uses TokenHarbor's free allowance and is not billed; availability, quotas and model access remain account-dependent.
Limits above: TokenHarbor. Available gateway routes are listed below.
https://tokenharbor.ai/models/mimo-v2.6-flash:free explicitly identifies Xiaomi's MiMo V2.6 Flash, matching the supplied reviewed OpenCode entry. The existing creator evidence at https://mimo.xiaomi.com/blog/mimo-v2-6-tool-call-repetition maps the MOPD release to the unchanged mimo-v2.6-flash API name. The shared duplicate_key denotes the advertised Flash model, without asserting a pinned TokenHarbor checkpoint or identical gateway settings.
DeepSeek's multimodal reasoning model adds image understanding to coding and text analysis. Its encoder-decoder design separates reading a long input from generating the answer, reducing the amount of information held in memory during generation.
Details & thinking
The exact route uses TokenHarbor's free allowance and is not billed; availability, quotas and model access remain account-dependent.
Limits above: TokenHarbor. Available gateway routes are listed below.
https://tokenharbor.ai/models/deepseek-v4.1-flash:free explicitly identifies DeepSeek V4.1 Flash, matching the supplied reviewed NVIDIA and Infron entries and creator card https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash. Shared advertised identity does not establish identical gateway settings or pinned checkpoint revisions.
An anonymous reasoning preview for coding and long-document analysis, with tool-calling support. The developer and underlying checkpoint have not been disclosed, so no model-family or parameter-count claims can be verified.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
The supplied Kilo catalog snapshot records stealth/glyph-cluster. Vercel's official listing at https://vercel.com/ai-gateway/models/glyph-cluster independently documents this public alias as an anonymous early-access reasoning model. Neither establishes its creator, checkpoint revision or exact checkpoint equivalence across gateways, so identity remains route-specific.
Upstage's compact model is designed for document extraction, classification and structured JSON output. It activates 3 billion of 35 billion parameters, with optional reasoning for requests that involve several steps or tool calls.
Details & thinking
The creator describes limited-time free Hermes Agent access starting October 5, 2026. Nous currently lists Solar Mini 4 as free; continued free availability is not established.
Limits above: Nous Portal. Available gateway routes are listed below.
The supplied Nous route upstage/solar-mini4:free and the official Nous listing identify Solar Mini 4 by Upstage, matching the creator's announcement and solar-mini4 Chat API alias. The gateway does not establish a pinned checkpoint revision. Solar Mini 4 is distinct from the known Solar Pro 4 entry; no known sibling requires identity merging or display grouping.
Liquid AI's compact text model is designed for local agents, tool calling, structured outputs and domain-specific fine-tuning. Its 1.73-billion-parameter architecture combines gated short convolutions with sliding-window and global attention.
Details & thinking
Cline's exact free-route catalog advertises zero input and output token prices. Separately, the supplied successful test at 2026-11-07T12:05:33+00:00 establishes access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: Cline. Available gateway routes are listed below.
Cline's official full catalog explicitly maps liquid/lfm-3-1.7b:free to LiquidAI/LFM3-1.7B. The creator card identifies the post-trained text checkpoint, separately from LFM3-1.7B-Base. No serving precision or repository revision is pinned.
A vision-language model for interpreting quantum processor calibration experiments, extracting measurements and explaining experimental results.
Details & thinking
Listed as a Free Endpoint on NVIDIA Build and successfully tested by the current scan. Availability and quotas may change.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA's exact route model card identifies this named checkpoint. Shared identity with other aggregators refers to the advertised model, not identical serving limits or quantization.
Meta's instruction-tuned vision-language model supports image captioning, visual questions and document understanding. It connects a separately trained vision adapter to a Llama 3.1 language model through cross-attention layers.
Details & thinking
The supplied successful test at 2026-11-05T18:53:05+00:00 establishes access to this exact NVIDIA route without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA's exact route reference identifies Meta's Llama-3.2-90B-Vision-Instruct, corresponding to the creator repository meta-llama/Llama-3.2-90B-Vision-Instruct. This identifies the instruction-tuned 90B vision release rather than its base or 11B variants. No pinned serving revision is established.
Meta's multimodal model supports coding, planning and reasoning over screenshots, charts and documents. Distilled from Muse Spark, it combines a dense language model with a dedicated perception encoder and is designed for deployment on consumer hardware.
Details & thinking
NVIDIA's exact route page advertises a free API endpoint. Separately, the supplied successful test at 2026-11-05T14:17:29+00:00 establishes access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA's exact muse-glimmer-30b listing links Meta's meta-models/Muse-Glimmer-30B repository. The disclosed creator is Meta; NVIDIA is the serving gateway. No pinned precision or repository revision is established.
A compact language model for responsive assistants and tool-using agents, with optional reasoning.
Details & thinking
Listed as a Free Endpoint on NVIDIA Build and successfully tested by the current scan. Availability and quotas may change.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA's exact route model card identifies this named checkpoint. Shared identity with other aggregators refers to the advertised model, not identical serving limits or quantization.
MiniMax's text model is designed for software engineering and multi-step productivity tasks using application-supplied tools. Its creator documents specialization in log analysis, debugging and refactoring, alongside repeated editing of documents, spreadsheets and presentations.
Details & thinking
The supplied exact-route catalog evidence marks usage_based_only as false while also advertising paid list prices. Access and quotas depend on the account. No valid inference output or model-specific top-up requirement is supplied; public aggregate statistics and example requests are not dedicated-account access tests.
Limits above: LLM7. Available gateway routes are listed below.
LLM7's exact-route page identifies minimax-m2.7 and attributes it to MiniMax. The creator's MiniMax-M2.7 repository establishes the advertised release and published weights. Preserved the existing duplicate_key; neither source pins LLM7's serving precision or repository revision. MiniMax M3 and M3.1 Flash remain distinct versions.
Intern Science's model is designed for extended research, engineering and tool-calling tasks. Its mixture-of-experts architecture has 35 billion parameters, with training that combines specialized teachers across multiple domains.
Details & thinking
Limits above: Intern Discovery. Available gateway routes are listed below.
The supplied Discovery catalog identifies Agents-A1 and the existing metadata attributes it to Intern Science. The exact creator repository https://huggingface.co/InternScience/Agents-A1 confirms that attribution and distinguishes the 35B model from Agents-A1-4B. No pinned serving revision is established.
Atria Dawn is designed for research analysis, coding and multi-turn tasks using application-supplied tools. It accepts text and can work with research excerpts, code and tool results to develop analyses and revise implementations.
Details & thinking
Limits above: Intern Discovery. Available gateway routes are listed below.
The supplied Discovery listing contains Atria-Dawn. The official Atria API documentation at https://api.atria-asi.ai/docs now explicitly documents that exact ID. The supplied Atria creator card establishes Shanghai Artificial Intelligence Laboratory's authorship. The documentation does not establish checkpoint equivalence with Atria-Dawn-Preview, so the identities remain separate.
A coding-focused model from Mistral for generating and completing source code. Its fill-in-the-middle training helps it continue code between existing blocks, making it useful for editor assistance and code-related conversations. The displayed 32K output figure is a shared-context upper bound: input tokens reduce the remaining output budget.
Details & thinking
The supplied exact-route catalog evidence marks usage_based_only as false while also advertising paid list prices. Separately, successful tests at 2026-11-05T13:01:41+00:00 and 2026-11-05T18:43:53+00:00 establish access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: LLM7. Available gateway routes are listed below.
LLM7's official exact-route page identifies codestral-latest and Mistral, but does not map this alias to a dated checkpoint. Preserved its separate duplicate_key. The supplied Codestral 25.08 page supports family capabilities rather than establishing this gateway alias's exact destination.
Cohere's coding specialist for editing repositories, generating code and working through terminal tasks. It has 30 billion parameters with roughly 3 billion active per token, and can reason between successive tool calls.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
Preserved the established exact-model identity shared with the supplied Cline and OpenRouter North Mini Code entries, supported by https://huggingface.co/CohereLabs/North-Mini-Code-1.0. Shared advertised identity does not establish identical gateway settings or pinned revisions.
The July 31 revision of DeepSeek V4 Flash is post-trained for coding and multi-step agent tasks. It also includes a speculative-decoding module, which lets compatible inference servers draft and verify tokens more efficiently.
Details & thinking
The exact route uses TokenHarbor's free allowance and is not billed; availability, quotas and model access remain account-dependent.
Limits above: TokenHarbor. Available gateway routes are listed below.
TokenHarbor explicitly states at https://www.tokenharbor.ai/blog/deepseek-v4-flash-0731-vs-pro-new-agent-benchmarks-explained that both deepseek-v4-flash and deepseek-v4-flash:free serve Flash 0731. This establishes the same advertised dated model as the supplied Infron 0731 entry and https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731. The unchanged runtime ID does not identify the original preview. Undated Infron and OpenCode identities are not merged.
Dots Studio's multimodal reasoning model can read documents and charts, interpret video and understand audio. Its sparse architecture activates 16 billion of 280 billion parameters; reasoning can be enabled for more involved questions.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
Google's multimodal reasoning preview supports coding and interactive tasks using application-supplied tools. It combines visual and audio understanding with text generation, and offers several thinking levels to adjust the depth of analysis.
Details & thinking
Google advertises free-tier input and output for this exact preview. Quotas and eligibility apply; no successful account-access test is supplied.
Limits above: Google AI Studio. Available gateway routes are listed below.
Google's lightweight multimodal model is designed for translation, transcription and information extraction at scale. It can produce structured JSON from documents or messages and apply additional reasoning to more involved requests.
Details & thinking
Google advertises free-tier input and output for the stable model. Quotas and eligibility apply; no successful account-access test is supplied.
Limits above: Google AI Studio. Available gateway routes are listed below.
Google's multimodal Flash-Lite preview generates text from language, visual and audio inputs. It supports structured responses, reasoning and function calls, allowing applications to process varied material and request actions from their own tools.
Details & thinking
Limits above: Google AI Studio. Available gateway routes are listed below.
Google's multimodal Flash-Lite model targets document parsing and high-volume tasks delegated by applications. It combines structured extraction and text generation with adjustable reasoning, using minimal thinking by default for lightweight work.
Details & thinking
Google advertises free-tier input and output for this exact model. Quotas and eligibility apply; no successful account-access test is supplied.
Limits above: Google AI Studio. Available gateway routes are listed below.
Google's Gemini 3.6 Flash combines multimodal understanding with code generation and spatial reasoning. It is designed for iterative coding and tool-based tasks where an application repeatedly evaluates results and requests the next action.
Details & thinking
Limits above: Google AI Studio. Available gateway routes are listed below.
Google explicitly states at https://ai.google.dev/gemini-api/docs/models/gemini-3.5-flash that API requests to gemini-3.5-flash are automatically routed to gemini-3.6-flash. The supplied models/gemini-3.5-flash upstream ID is preserved, while the name and duplicate_key identify the documented current destination. This is a sourced alias mapping, not an inference from version similarity; it does not pin a checkpoint revision.
Google's vision-language model supports spatial reasoning and planning for robotics applications. It interprets visual scenes and can coordinate multi-step tasks through application-supplied tools, including workflows involving several robots.
Details & thinking
Limits above: Google AI Studio. Available gateway routes are listed below.
Google's sparse Gemma 4 model combines image understanding with multilingual writing, coding and reasoning. It activates roughly 4 billion of its 26 billion parameters per token, making it the lighter-compute alternative to the dense 31B version.
Details & thinking
Limits above: Google AI Studio. Available gateway routes are listed below.
Google's hosted Gemma guide explicitly lists gemma-4-26b-a4b-it. This matches the exact instruction-model version in the supplied reviewed Cline and OpenRouter entries and https://huggingface.co/google/gemma-4-26B-A4B-it. Shared advertised identity does not establish identical gateway settings or pinned weight revisions.
Google's dense 31-billion-parameter Gemma 4 model handles text and images, with optional thinking for harder questions. It supports multilingual writing, visual analysis, coding and function calls for applications that supply their own tools.
Details & thinking
NVIDIA's exact route page advertises a free API endpoint. Separately, the supplied successful test at 2026-11-05T18:45:35+00:00 establishes access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA's exact route explicitly links Google's gemma-4-31B-it instruction checkpoint, establishing the advertised identity already shared by the supplied Google, Cline, Kilo and OpenRouter entries. Serving settings and pinned revisions are not assumed identical.
Z.ai's text reasoning model focuses on complex software engineering, extended tool use and cybersecurity tasks. It builds on GLM 5.2 with further training for coding and long-running work, rather than a new base architecture.
Details & thinking
Limits above: Intern Discovery. Available gateway routes are listed below.
The supplied Discovery glm-5.3 listing matches the exact version in the reviewed NVIDIA z-ai/glm-5.3 entry and its creator evidence at https://huggingface.co/zai-org/GLM-5.3. Reused the established advertised-model identity without assuming identical serving settings or pinned weight revisions.
Z.ai's multimodal Flash model combines visual understanding with coding and tool use. Its sparse architecture activates 18 billion of 320 billion parameters and mixes attention mechanisms to handle long inputs efficiently.
Details & thinking
The supplied October 14 test and recorded successful tests from November 5 through November 12 establish access to this exact River route without a top-up using the dedicated account that had never held credit or linked a payment method. The latest supplied successful test at 2026-11-13T12:03:08+00:00 establishes the same access at that time. Separately, the supplied prior review records advertised prices of $0.12 input, $0.024 cached input and $0.40 output per million tokens. Current documentation requires JavaScript to display prices. Continued availability and quotas remain unverified.
Limits above: River. Available gateway routes are listed below.
River explicitly lists zai-org/GLM-5.3-Flash at https://docs.river.ai/guides/inference-models/, matching the creator repository and supplied reviewed standard Flash entries. Preserved the established duplicate_key and exact runtime namespace. Shared advertised identity does not establish identical serving configurations or pinned revisions.
OpenAI's larger open-weight reasoning model is designed for analysis, coding and tool use. Three reasoning levels let applications trade speed for deeper analysis; the weights can run on a single GPU with 80 GB of memory.
Details & thinking
Limits above: Groq. Available gateway routes are listed below.
OpenAI's smaller open-weight reasoning model supports coding, text analysis and function calls, with three reasoning levels. Its compact memory footprint allows local deployment with about 16 GB of memory, unlike the much larger 120B version.
Details & thinking
Limits above: Groq. Available gateway routes are listed below.
Intern's 397-billion-parameter multimodal model targets scientific reasoning and extended tasks using tools. Its visual pretraining learns directly from scientific literature pages, preserving relationships between text, symbols and visual information.
Details & thinking
Limits above: Intern Discovery. Available gateway routes are listed below.
Preserved the supplied Intern Large Models creator. Its official model documentation identifies intern-s2 as the released 397B model, separately from Intern-S2-Preview-397B and the 35B preview. The corresponding creator repository is https://huggingface.co/internlm/Intern-S2-397B. Discovery's supplied exact ID identifies the advertised release without pinning its serving revision.
A language model for coding assistants and agent workflows that interleave reasoning with tool calls.
Details & thinking
Listed as a Free Endpoint on NVIDIA Build and successfully tested by the current scan. Availability and quotas may change.
Limits above: NVIDIA. Available gateway routes are listed below.
NVIDIA's exact route model card identifies this named checkpoint. Shared identity with other aggregators refers to the advertised model, not identical serving limits or quantization.
Liquid AI's small reasoning model is designed for on-device use, including document extraction, retrieval-based answers and tool calls. Its 2.6-billion-parameter hybrid architecture targets memory-constrained hardware rather than large-scale coding tasks.
Details & thinking
Limits above: OpenRouter. Available gateway routes are listed below.
A 12-billion-parameter instruction-tuned model developed by Mistral and NVIDIA for multilingual conversation, coding and document tasks. Its 128K context window supports long inputs, and its weights are available under Apache 2.0. The displayed 128K output figure is a shared-context upper bound: input tokens reduce the remaining output budget.
Details & thinking
The supplied exact-route catalog evidence marks usage_based_only as false while also advertising paid list prices. Separately, successful tests at 2026-11-05T13:02:06+00:00, 2026-11-05T14:11:58+00:00 and 2026-11-05T18:44:16+00:00 establish access without a top-up using the dedicated account that had never held credit or linked a payment method. Continued availability and quotas are not established.
Limits above: LLM7. Available gateway routes are listed below.
The dated upstream ID matches the official mistralai/Mistral-Nemo-Instruct-2407 checkpoint, jointly released by Mistral and NVIDIA in July 2024.
Google's rolling Flash-Lite alias represents the lightweight Gemini model variation for text generation. Its current version is unresolved, so specific specializations, multimodal capabilities and reasoning controls cannot be assigned from a presumed destination.
Details & thinking
Limits above: Google AI Studio. Available gateway routes are listed below.
The supplied Google catalog snapshot identifies this exact alias. The rechecked official model catalog and release notes do not explicitly establish its current destination. Latest-alias semantics alone do not identify a version; known 2.5, 3.1 and 3.5 Flash-Lite releases remain separate.
NVIDIA's multimodal reasoning model brings text, images, audio and video into one conversation. It is designed for tasks such as reading charts and documents, transcribing speech and understanding what happens in a video.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
NVIDIA's largest Nemotron 3 model targets extended coding, scientific reasoning and multi-step tool use. Its sparse architecture activates 55 billion of 550 billion parameters, with an emphasis on retaining information across long tasks.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
Preserved the established exact-model identity shared with the supplied Cline, OpenRouter and explicitly mapped OpenCode Ultra entries, supported by https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16. Shared advertised identity does not establish identical serving configurations or pinned revisions.
NVIDIA's compact reasoning model uses a sparse architecture with 30 billion total parameters and 3 billion active per token. Its design combines long-context processing with inference optimizations for applications where response time matters.
Details & thinking
Limits above: Kilo. Available gateway routes are listed below.
Alibaba's dense 27-billion-parameter model combines coding and reasoning with image and video understanding. It can analyze visual material alongside text and use application-supplied tools while working through longer tasks.
Details & thinking
Limits above: Groq. Available gateway routes are listed below.
An LLM API lets an application send a request to a language model over HTTP and receive a response. You normally need a base URL, an API key, and the exact model ID; an OpenAI-compatible client can then make the request.
What is the difference between a model provider and an API aggregator?
A provider develops or operates the underlying model. An API aggregator exposes models from one or more providers through its own account, endpoint, pricing, and usage rules, so the same model may appear through several different routes.
LiteLLM is an open-source proxy that presents an OpenAI-style interface in front of many model providers and aggregators. You do not need LiteLLM to call a route directly; it is useful when you want one configuration and endpoint for several backends. The GitHub importer keeps that route list in sync.
Where can I use these models?
Every tool below needs the same three values — API base URL, API key, and model ID — only the settings screen differs.
Linux
Windows
macOS
All tools below run on Linux, Windows, and macOS.
OpenClaw
Use an assistant for files, messages, and everyday tasks.
A chat window answers questions. An agent does things: it opens a browser, reads files, calls other tools and reports back. The model is the brain; the API key is what lets your agent borrow it.
Browser automation. Tell OpenClaw “check whether my parcel has shipped and send me the status” — it opens the carrier’s page, reads the tracking details and messages you on Telegram.
Tools via MCP. Hermes Agent connected to your calendar, Google Drive or home server can compare your week, move a file or restart a service when you ask in plain words.
Working in your files. Open a project folder in OpenCode and ask it to fix a failing script or explain an unfamiliar config — it reads the code, makes the change and shows you what it did.
A free route is enough for all of this. When one aggregator is busy, LiteLLM switches to the next.
Twice a day, we discover models directly from the aggregators' own API catalogs, such as /v1/models, and send real requests to the selected free model routes. We use API keys from dedicated test accounts that have never held any account credit and have never had a payment method linked: no credit card, PayPal or cryptocurrency payments. A successful HTTP status alone is not enough: the response must contain model output. A passing test confirms access with these unfunded accounts at the time of testing. We test each model and aggregator combination separately. Routes that explicitly require OpenCode or an agentic harness receive a separate test through OpenCode and are marked accordingly. Only successful routes are published. Account and usage conditions can vary by route and change over time. We regularly add aggregators to expand coverage; each scan discovers new models within the configured catalogs.
What happens when I click “Try this model here and now”?
The browser sends only a fixed route, question, and emoji selection. The server builds the request, applies its configured safety checks, sends it to the selected backend, validates the response, and returns the result. The demo shares one site-wide budget of 30 test attempts per UTC clock hour; failed upstream attempts count too.
Are these models really free, and what limits apply?
Every model and aggregator combination listed here passed a real test with an account that has never held credit or had a payment method linked. No payment or top-up was needed. Context and output limits appear on the model cards; request limits depend on the aggregator, model and account. Availability and limits can change.
Why do aggregators ask for a credit card if it is not required?
Aggregators offer paid models and higher quotas alongside their free offerings. Their dashboards may ask for payment details for those other offerings. Every model and aggregator combination listed in this directory was successfully tested without account credit or a linked payment method. Payment requirements for other models do not apply to the access verified here.
