KageMux chatModel notices and use conditions
Built with Llama. Routes use the hosting provider shown in the picker, including Cloudflare and optional AI Horde volunteer workers. KageMux does not distribute model weights or claim ownership of supplier models.
Llama 3.1 and Llama 3.2 are licensed under their respective Community License Agreements, Copyright © Meta Platforms, Inc. All Rights Reserved. Their acceptable-use policies apply. Gemma and derivatives are subject to the Gemma Terms of Use and Prohibited Use Policy. NVIDIA Nemotron and derived PhoneLLM models remain subject to the NVIDIA Nemotron Open Model License; PhoneLLM modifications additionally use BSD 2-Clause.
Gemma Terms of Use · Gemma prohibited uses · Llama 3.1 acceptable use · Llama 3.2 acceptable use · Apache License 2.0
Mistral 7B Instruct V0.2 LoRA · Cloudflare
cf/@cf/mistral/mistral-7b-instruct-v0.2-lora
License: apache-2.0. Publisher: mistralai/Mistral-7B-Instruct-v0.2.
Publisher model card · Publisher terms
Mistral Small 3.1 24B · Cloudflare
cf/@cf/mistralai/mistral-small-3.1-24b-instruct
License: apache-2.0. Publisher: mistralai/Mistral-Small-3.1-24B-Instruct-2503.
Publisher model card · Publisher terms
Nemotron 3 Super 120B A12B · Cloudflare
cf/@cf/nvidia/nemotron-3-120b-a12b
License: nvidia-nemotron-open-model-license. Publisher: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8.
Publisher model card · Publisher terms
Model limits and provider allowance are shared across the service. Availability can change; a directory listing is not access.
Local lightweight model notices
Qwen 2.5 0.5B, Qwen 3 0.6B and SmolLM2 135M use Apache License 2.0. KageMux runs them locally; these small models are not equivalent to larger hosted models. See Qwen 2.5, Qwen 3, SmolLM2 and the Apache license copy.
Powered by UncloseAI
The Qwen3.8-27B-exl3 community route is served by UncloseAI, not by KageMux-owned GPUs. Its terms, privacy policy and the model publisher notices apply. KageMux does not claim endorsement by that provider.
Cloudflare structured-analysis models
Clef, Clef Flash and Clef Omni use publisher Apache 2.0 licenses. They are hosted APIs; KageMux does not distribute their model weights. See the publisher licenses for Clef, Clef Flash and Clef Omni. Clairic is available under Cloudflare preview terms. These are structured tools, not free-form conversational models; usage instructions describe supported inputs and limits.
Optional AI Horde volunteer inference
AI Horde requests are served by independent volunteer workers. Your recent conversation text and pinned note are sent only after explicit model selection and the separate processing consent. Do not send sensitive information. Confidentiality and availability are not guaranteed. Auto never chooses this provider. Replies are limited to 128 generated tokens; reported resource usage is preserved in kudos, not converted into token totals. No provider account or credits are created for each customer.
AI Horde privacy policy · AI Horde terms · Volunteer API limitations. Each underlying model remains subject to its publisher terms: Apache 2.0 for Gemma 4 and the Gemma Terms of Use for Gemma 3. KageMux does not distribute model weights.
Additional AI Horde text routes
Built with Llama. Llama 3.2 3B Q4 is subject to the Llama 3.2 Community License and acceptable-use conditions. Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
This route was verified for short text replies only; no image, audio or large-context support is advertised. Provider-reported model identifiers are checked; volunteer hardware and weights are not independently audited. Llama 3.2 license copy.
SmolLM2 360M Instruct — dedicated local llama.cpp
Publisher: HuggingFaceTB. Model: SmolLM2-360M-Instruct-GGUF, official Q8_0 weights at revision 593b5a2e04c8f3e4ee880263f93e0bd2901ad47f. License: Apache 2.0. The llama.cpp runtime is MIT-licensed; model weights are not distributed by this website.
This is a lightweight English text model, not a frontier reasoning model. It runs on the operator computer using a dedicated loopback-only CPU service; the browser does not connect to that computer directly. Maximum output is 128 tokens, configured context 4,096 tokens, with one shared generation slot. No tools, images or audio are enabled. Sign-in, provider-processing consent and normal account limits still apply. It is available by explicit model selection only; Auto never uses this local route.
Llama 3.2 1B Instruct — optional AI Horde route
Built with Llama. Publisher: Meta. Publisher model card; Llama 3.2 Community License. Copyright © Meta Platforms, Inc. All Rights Reserved.
This exact hosted route returned the expected model identity and remembered synthetic conversation context. It is supplied by independent volunteer workers, not Meta hosting. Separate volunteer-processing consent is mandatory. Do not send sensitive data. Maximum reply is 128 tokens; capacity is shared, variable and low-priority. Usage is recorded in provider-native kudos, not invented token counts. This route is never selected by Auto.
SmolLM2 1.7B Instruct Q4 — local CPU chat
Publisher: HuggingFaceTB. Publisher model card; Apache-2.0 license. Hosted locally using Ollama, source tag smollm2:1.7b-instruct-q4_K_M. This is the 1.7 billion parameter model, not a 17B or frontier model.
Explicit selection only, never selected by Auto. Replies capped at 128 tokens, context configured at 4096 tokens subject to KageMux transcript limits. One shared request at a time. Requires the operator computer to be online. Real local gateway tests passed reported model identity, short conversation recall and a basic instruction check; no claim of a signed-in customer or general reasoning benchmark.
Optional public inference providers
These providers are never chosen by Auto. You must explicitly select one and accept its processing notice before recent saved messages and your pinned note are sent. Do not send confidential, regulated or sensitive personal data.
Powered by UncloseAI
Qwen 3.8 27B is processed by the UncloseAI public service in the United States. Requests may be scanned and randomly captured for review; access logs may be retained for up to one year. Shared KageMux limit: one generation per minute. Privacy · Terms.
Powered by Pollinations Classic and OVH
This route supplies GPT-OSS 20B, not proprietary ChatGPT. Pollinations processes requests and may keep generated-response caches for up to 30 days. Downstream provider policies apply; no no-retention guarantee is made. Pollinations says it does not train on prompts or outputs without opt-in. Shared rate limit: one request at a time and at least 30 seconds apart. We do not pay or retry automatically when its free slot is exhausted. The supplier service is for users aged 16+. Privacy · Terms · GPT-OSS Apache 2.0 license.
Both services are best-effort, with shared capacity and no uptime promise. Declining processing consent leaves your conversation saved without calling that provider.
Qwen 3.5 122B A10B — AI Horde
This is the Qwen/Qwen3.5-122B-A10B model identified by an AI Horde volunteer worker, not proprietary ChatGPT or Gemini. Text chat only in KageMux. Replies use the existing 128-token output limit and resource usage is reported in kudos. Volunteer availability and latency can change; this is not guaranteed hosted capacity. Explicit AI Horde processing consent is required and Auto never selects it.
Publisher: Qwen / Alibaba. License: Apache 2.0. Publisher model card and notices. KageMux does not distribute model weights. Do not send confidential or sensitive personal data to volunteer workers.
Translation, sentiment and passage-ranking tools
These are dedicated text-task models, not conversational assistants. Only the current explicit task input is sent to Cloudflare. Translation and reranking retain the usage numbers supplied by the provider. The sentiment endpoint does not report tokens, so only an observed completed-request count is recorded.
cf/@cf/meta/m2m100-1.2b — mit; publisher model card and license.
cf/@cf/huggingface/distilbert-sst-2-int8 — apache-2.0; publisher model card and license.
cf/@cf/baai/bge-reranker-base — mit; publisher model card and license.
Sentiment classifications can reflect dataset bias and do not establish facts about people. Reranking scores are not calibrated probabilities. Input guide and limits.