# Moore Threads (摩尔线程) > Moore Threads is a Chinese GPU company founded in June 2020, building full-function GPUs and the > MUSA (Meta-computing Unified System Architecture) software stack — a CUDA-alternative > programming environment. Alongside the silicon it operates a small public API surface: the > AIBook speech platform (text-to-speech, speech recognition, voice conversion), the KUAE Cloud > Coding Plan LLM inference endpoint, and a digital-human Web SDK. This file was GENERATED by API Evangelist from Moore Threads' own public documentation on 2026-08-26. Moore Threads does not publish an llms.txt (probed https://www.mthreads.com/llms.txt 404, https://developer.mthreads.com/llms.txt 404, https://docs.mthreads.com/llms.txt returns the Docusaurus HTML shell). It is not an official Moore Threads document. No OpenAPI, AsyncAPI, GraphQL SDL, MCP server or A2A agent card is published on any Moore Threads host. Every operation below is documented in prose and parameter tables only. ## APIs - [AIBook Text-to-Speech — HTTP](https://docs.mthreads.com/tts/unary-tts/introduction): POST https://aibook-api.mthreads.com:32414/api/v1/tts/generate. Up to 300 UTF-8 characters to wav or mp3. Parameters: text, voice, format, volume (0-100), speech_rate (-500..500). Authorization header carries the access token. 27 published voices, split between large-model (`_llm` suffix) and lite voices. - [AIBook Text-to-Speech — streaming WebSocket](https://docs.mthreads.com/tts/stream-tts/introduction): wss://aibook-api.mthreads.com:32414/api/v2/tts/stream_generate. Returns base64 PCM chunks with optional word-level (enable_subtitle) and phoneme-level (enable_phoneme_timestamp) timestamps. Also accepts a STREAMING TEXT input so an LLM's tokens can be voiced as they are produced, with prosody coherence maintained across fragments. One synthesis per connection. Clusters: mt_tts, mt_llm_tts. - [AIBook Speech Recognition — realtime WebSocket](https://docs.mthreads.com/asr/realtime-recognition/introduction): wss://aibook-api.mthreads.com:62220/api/v1/asr?token=... 16 kHz 16-bit mono PCM (also OPUS, AMR, MP3, AAC). Flow: StartTranscription -> binary frames -> StopTranscription. Events: SentenceBegin, SentenceChanged, SentenceEnd, TranscriptionCompleted. Options include punctuation, ITN, disfluency removal, speaker diarization, nbest, confidence, word-level timing and semantic sentence detection. Send audio at least every 10 seconds or the session ends. - [AIBook Speech Recognition — recording files](https://docs.mthreads.com/asr/recording-recognition/introduction): https://aibook-api.mthreads.com:62220/api/v1/asr with Authorization: Bearer. Submit by URL or by chunked upload, then poll (Query) or receive the result on a `callback` URL. Audio to 512 MB / 5 hours; results retained 72 hours and retrievable up to 100 times. - [AIBook Streaming Voice Conversion](https://docs.mthreads.com/vc/stream-vc/introduction): wss://aibook-api.mthreads.com:32314/api/v1/streaming_vc?token=... 16 kHz PCM in, 48 kHz PCM out. Flow: StartConversion -> binary frames -> StopConversion. - [KUAE Cloud Coding Plan](https://docs.mthreads.com/kuaecloud/kuaecloud-doc-online/coding_plan/tools_config): LLM inference on Moore Threads MTT S5000 GPUs, serving GLM-4.7. Speaks two protocols from one host — the Anthropic Messages protocol at https://coding-plan-endpoint.kuaecloud.net (Claude Code) and the OpenAI Chat Completions protocol at https://coding-plan-endpoint.kuaecloud.net/v1 (Cline, Cursor, Roo Code, Kilo Code, OpenCode). ## Authentication - AIBook speech APIs: opaque access token, issued only by contacting Moore Threads staff. No signup flow. Header on HTTP surfaces, `?token=` query parameter on WebSocket surfaces. - KUAE Cloud Coding Plan: self-service API key created at https://coding-plan.kuaecloud.net/apikey after claiming the free trial or redeeming a purchased code. - No OAuth 2.0, no OpenID Connect, no scopes, no mTLS anywhere. ## Limits - Coding Plan quota is a rolling 5-hour prompt budget: Free Trial ~40, Lite ~120, Pro ~600, Max ~2400. No overage — wait for the next window. - TTS: 300 characters per request. - ASR: 512 MB / 5 hours per file; 72-hour result retention; 100 result fetches; 10 MB upload chunks. - No RateLimit-*, X-RateLimit-* or Retry-After header on any surface. Exhaustion is signalled in-band by status 2003 (too many requests) and 2004 (service overload). ## Errors All AIBook surfaces return HTTP 200 with a numeric `status` and `status_text` in the body. 1xxx success/in-progress, 2xxx identity and quota, 3xxx server, 4xxx request validation. Not RFC 9457. The KUAE Cloud endpoint returns the Anthropic/OpenAI error envelopes. ## Reversibility Read-only in effect. Nothing on the public surface creates durable customer state, and no cancel, undo or delete operation is documented for a submitted ASR task. ## SDKs - `mtai` (npm, TypeScript, v0.2.21, published 2026-07-23) — digital-human Web SDK for the 摩影元像 product. Targets a customer-hosted server, not a Moore Threads endpoint. - No client library exists for any AIBook speech API in any registry. - Docker Hub: mthreads/musa, mthreads/ollama, mthreads/vulkan-sdk. ## Software stack (not APIs) - [MUSA SDK](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/introduction) — compiler toolchain (mcc), runtime, muDNN, muBLAS, muFFT, muSPARSE, muSOLVER, MCCL, Triton-MUSA, TileLang-MUSA, Musify CUDA porting. Current release 5.2.0. - [GitHub organisation](https://github.com/MooreThreads) — 133 public repositories, largely MUSA ports of upstream CUDA projects (torch_musa, vllm-musa, ollama-musa, tilelang_musa, mutlass). ## Links - Website: https://www.mthreads.com/ (English: https://en.mthreads.com/) - Documentation: https://docs.mthreads.com/ - Developer portal: https://developer.mthreads.com - Blog: https://blog.mthreads.com - Support / FAQ: https://www.mthreads.com/support/FAQ - Terms: https://docs.mthreads.com/kuaecloud/kuaecloud-doc-online/terms/user_agreement - Privacy: https://www.mthreads.com/privacy