Muse Spark Open vs Tencent Cloud TokenHub

Muse Spark Open 44

Read as software the owner runs. No account or key to run it. Meta's weights need manual approval on Hugging Face and an access token, and server authentication is the owner's to add (12 of 30). An agent model that calls whatever tools the operator gives it. The card recommends human-in-the-loop confirmation for irreversible actions and Meta reports agentic-risk evaluations, but nothing in the model or template enforces an approval (8 of 20). The chat template neutralises special-token sequences and media placeholders in caller text and wraps tool results in <tool_output>, and the card advises protecting internal assistant messages and layering defences against adversarial input. No indirect prompt-injection results are published (10 of 15). No call log of its own. Visibility depends on the serving stack (5 of 15). Weights are safetensors that load through Transformers' own classes from 5.19, and Meta published cyber, chemical, biological and misalignment evaluations. No security.txt on dev.meta.ai or research.meta.ai, and no security route on the card or in the docs, only museoss@meta.com and the Hugging Face discussions (9 of 20). 44.

Tencent Cloud TokenHub 71

Router, read on the model lines. Bearer keys can be limited to chosen models and inference services, up to 200 source IP ranges and daily, monthly or total token quotas, can be disabled or deleted, and can be created in code through CreateApiKey with signed Tencent Cloud credentials. Keys are long-lived with no expiry setting or built-in rotation (26 of 30). Clause 2.17 of the terms says inputs are not used to train Tencent Cloud's own models, and the FAQ says data is not used for training. Inputs to third-party deployed models (DeepSeek's directly supplied ones) go to DeepSeek under its terms (14 of 20). Clause 2.14 keeps data only as long as the service needs, with no periods stated. When a customer asks for generated content without AI labels, Tencent Cloud keeps logs of who received it for at least six months. No zero-retention option was found (7 of 15). Per-call audit logs (request ID, key, client IP, status, tokens, latency) and optional full call logs go to the customer's CLS, billed separately, with usage by key and a monitoring dashboard (14 of 15). No security.txt on cloud.tencent.com, www.tencent.com or tokenhub.tencentmaas.com (404). Tencent's bug bounty through TSRC lists Tencent Cloud in scope. No certification for TokenHub was checked (10 of 20).

Muse Spark Open 50

Self-hosted rule. Nothing to buy from Meta, and the licence grants use free of charge. The Model as a Service clause asks for a separate agreement only from licensee groups above $20M gross revenue in 12 months (20). Free to run with no card (20). Meta's own weights need a manually reviewed form with legal name and date of birth, so an agent can't get them alone. The ungated GGUF quantisations from ggml-org and Unsloth that Meta's docs point to download without approval, so part of the line (10 of 20). No payment protocol (0). Hosted copies at $1.25 input and $4.25 output per million tokens need each provider's own account and are not graded. 50.

Tencent Cloud TokenHub 35

No x402, MPP or L402 (0). Per-token prices for every model are public in CNY for Guangzhou and Singapore, with DeepSeek's peak and off-peak rates (20). Each account gets 1M free tokens per language model, valid for a year, in a campaign running to 31 December 2026. No card is mentioned, but the account needs real-name verification, so this line takes 15 of 20. A person signs up and completes real-name verification in a browser. Keys can then be created in code or by thcli auth login, but only from that account (0).

Muse Spark Open 69

Model reading of the checklist (tool use, structured output, caching, context, throughput, defaults and SDKs, errors). OpenAI-format tools and parallel calls come back as tool_calls, and the template supports namespaced tools and deferred loading through a tool_search tool. A forced tool_choice is not documented (17 of 20). JSON mode and JSON Schema come from each runtime's guided decoding, which Meta's docs show for TensorRT-LLM only (8 of 15). Prefix caching is the runtime's. Meta's SGLang command turns on hierarchical caching, and Dynamo routes each request to the worker holding its cached context (8 of 10). 1,048,576-token context, with sliding-window attention on 46 of 62 layers holding KV memory down (15 of 15). Draft models for speculative decoding in SGLang and vLLM, and no batch mode of its own (7 of 10). Any OpenAI SDK works against a local server, but reasoning can't be turned off, defaults to the most expensive level, and needs streaming, a max_tokens budget for reasoning and the right stop tokens (6 of 15). The template raises named errors for an unsupported reasoning effort, tool arguments passed as a string and a tool message with no name. Server errors are the runtime's (8 of 15). 69.

Tencent Cloud TokenHub 83

Router, read on the model lines. Function calling is marked on most chat models in the model list, with behaviour depending on the model routed (14 of 20). Structured output is marked per model and missing on the MiniMax and Qwen models, and an unsupported response_format returns code 400006 (12 of 15). Cache-hit prices are published per model, Kimi K3.1 Pro lists five-minute and one-hour cache writes, and a guide covers prompt_cache_key and the X-Session-ID header for instance affinity (14 of 15). Most listed chat models have contexts of one million tokens, a judgement call since a router has no default model (15). Batch at half price covers only Hy3 and Hy4 preview, reads and writes JSONL in the customer's own COS bucket and is documented for the Guangzhou endpoint (5 of 10). No TokenHub inference SDK. The OpenAI and Anthropic SDKs work with a changed base URL, and Tencent Cloud SDK 3.0 and the thcli CLI cover the control plane (8 of 10). Errors return a code, English and Chinese messages, the source, the upstream code and status and a request ID, and stream errors arrive as SSE error frames (15 of 15).

Muse Spark Open 60

Read for an open-weight model released today. Released on 15 October 2026 (30). One release, BF16 and NVFP4 on the same day, with no history of revisions, successors or notice periods, and Meta Model API has no deprecation policy either (6 of 20). Three Hugging Face discussions opened on release day had no reply from Meta when read. Feedback goes to the discussions or museoss@meta.com, and Meta's docs and developer blog carry dated posts (8 of 25). Day-one support in vLLM, SGLang, TensorRT-LLM, Transformers, llama.cpp, MLX and Unsloth with Meta's own recipes, though SGLang's is a dev image and TensorRT-LLM's a feature branch (12 of 15). Current dependencies (Transformers 5.19), but no CI visible to us (4 of 10). 60.

Tencent Cloud TokenHub 73

Closed service, read on the model lines. The newest release note is dated 24 October 2026 and the price page was updated on 31 October (30). The four retirement notices read gave 21 to 33 days, and eligible callers are switched to a successor at its price. No minimum notice period is written, and DeepSeek-V4.1-Flash (directly supplied) was moved to the 1017 version under the same ids on 17 October 2026 in a notice dated two days later (7 of 12). 18 retirement notices from 21 April to 22 October 2026, 12 of them since 3 August (3 of 8). Dated release notes and announcements, ticket and phone support around the clock and a Q&A community (13 of 15). Replies on the THCLI repository were not sampled (3 of 10). Tencent Cloud SDK 3.0 in several languages covers the control plane and tencent-tokenhub-cli 1.0.2 was published on 30 October 2026. No inference SDK of its own (12 of 15). The CLI has four releases since 15 September 2026, lists no runtime dependencies and needs Node.js 18 or later. CI not checked (5 of 10).

Muse Spark Open 61

Read for a model the owner serves. No specification of its own. The contract is the Jinja chat template, which renders OpenAI-format tools with additionalProperties: false added and rejects bad values, plus the OpenAI-compatible servers in vLLM and SGLang with the muse parsers (10 of 25). dev.meta.ai/llms.txt lists eight Muse Spark Open pages, each with a Markdown copy (10). The card lists seven intended uses and six limitations, the docs say when to pick BF16 or NVFP4 and which reasoning strength suits which task, and the Usage Policy states what is out of scope (17 of 20). reasoning_effort is an enum of four values and anything else returns an error, and tools take JSON Schema. Structured output is documented only for TensorRT-LLM's response_format, not in the prompting guide or for vLLM and SGLang (11 of 15). curl and Python examples for every runtime and hosted provider, the rendered prompt for each case, and the template's own error messages, but no error reference for a served model (11 of 15). One release, with no changelog or version scheme beyond the licence's version line, Muse Spark Open (2 of 15). 61.

Tencent Cloud TokenHub 73

Router, scored on the API lines. No OpenAPI file for the inference routes. Field pages document the OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats as TokenHub implements them, and the control plane is a Tencent Cloud API 3.0 service (version 2026-03-22) with typed parameters, API Explorer and SDKs (12 of 25). An llms.txt at cloud.tencent.com/document/product/1823/llms.txt links a Markdown copy of each docs page (10). The docs say which models support thinking, tools, structured output and caching, how TokenHub departs from each standard protocol, and which DeepSeek model suits which job. The quickstart's examples still use deepseek-v3, which is not in the model list, and Qwen3.5 stays in the list with a past shutdown date (13 of 20). Protocol pages list each field with type, required flag and default, and control-plane parameters carry enums and required flags (10 of 15). Examples in curl, Python, Node.js, Java and Go, and an error page listing 30 codes with HTTP status, cause and fix (14 of 15). Paths carry /v1, release notes are dated monthly from 7 April to 24 October 2026, and the control-plane API has a dated history of 26 releases to 18 September 2026 (14 of 15).