MagmaRouter vs Fireworks AI
Fireworks AI is a large inference and fine-tuning platform with serverless per-token models, dedicated GPUs, and both OpenAI and Anthropic compatible endpoints. Its terms restrict sexually suggestive and objectionable input, which is where MagmaRouter's uncensored line is a different product.
Facts checked October 2026 against Fireworks AI’s own site, docs and policies.
| MagmaRouter | Fireworks AI | |
|---|---|---|
| What it is | Hosted gateway for one uncensored, tool capable model line (Xploded). | A generative AI inference platform offering serverless models, dedicated GPU deployments, custom model and LoRA uploads, and fine-tuning. |
| Pricing | Pay as you go, $5 minimum top-up, no subscription required. | Serverless is per token on prepaid credits (for example GLM 5.3 Flash at $0.15 input, $0.03 cached, $0.50 output per 1M tokens). Batch inference is 50% of serverless pricing. Dedicated deployments are per GPU second, $8.00 per hour for H100 and H200 up to $20.00 per hour for GB300. Fine-tuning is $0.50 to $40.00 per 1M training tokens. Minimum payment and free credits not documented. |
| OpenAI-compatible | Yes | Yes, OpenAI-compatible at https://api.fireworks.ai/inference/v1, including tool calls. If max_tokens would exceed the context window it is adjusted automatically rather than erroring. |
| Anthropic-compatible | Yes, /v1/messages (verified with Claude Code) | Yes, POST /v1/messages with streaming at https://api.fireworks.ai/inference, documented for Claude Code. count_tokens, batch through this endpoint and server-side tools are not supported, and model names must be Fireworks identifiers. |
| Uncensored / abliterated | Five uncensored models, the Xploded line: GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash, MiMo V2.6 Flash and Qwen3.8 27B. Every one calls tools. | No uncensored, abliterated or roleplay models in the serverless catalogue. The terms say input must not include nudity or other sexually suggestive content, vulgar or obscene content, or content they judge harmful or objectionable in their sole discretion. Your own models or LoRAs can be uploaded, on dedicated deployments only. |
| Prompt retention | No prompt or completion stored; model, tokens and cost kept 90 days; signed prompt-hash receipt | The privacy policy says prompt and generation data for open models is not logged or stored without explicit opt-in, and prompts, training data and API inputs are not used to train their models without explicit opt-in. |
| Per-key spend caps | Lifetime and monthly per-key caps, enforced by reserving the worst case and clamping max_tokens | 6,000 requests per minute account-wide with a payment method and credits, 10 RPM without. An account monthly spend limit, plus tiered maximum monthly spend ($50 to $50,000) on a page labelled Legacy Postpaid. Per-API-key budgets are not documented. |
Where Fireworks AI is the better choice
- Much broader platform: hundreds of serverless models, dedicated GPUs billed per second, custom model and LoRA uploads, fine-tuning (SFT and DPO) and 50% off batch inference, none of which MagmaRouter offers.
- Documented zero retention for open models unless you opt in, high account rate limits (6,000 RPM), and their own inference hardware.
Where MagmaRouter is different
- MagmaRouter's models are all uncensored (the Xploded line: GLM 5.3 Xploded, GLM 5.3 Flash Xploded, DeepSeek V4.1 Flash Xploded, MiMo V2.6 Flash Xploded, Qwen3.8 27B Xploded) and every one can do tool calling for Claude Code, Cline and OpenCode, while Fireworks' terms restrict sexually suggestive and objectionable input.
- MagmaRouter enforces hard lifetime and monthly caps per key by reserving the worst-case cost and clamping max_tokens before each call, pay as you go from a $5 minimum top-up with no subscription.
The uncensored models MagmaRouter sells
If you are comparing Fireworks AI because you want an uncensored or abliterated model API, these are the models on MagmaRouter: GLM 5.3 uncensored, GLM 5.3 Flash uncensored, DeepSeek V4.1 Flash uncensored, MiMo V2.6 Flash abliterated, Qwen3.8 27B uncensored and Qwen3.8 27B Fast uncensored. Every one calls tools, works in Claude Code through the Anthropic endpoint, and is priced per million tokens as listed on its model page.
| Model | What it is | Context | Price in / out | Supports |
|---|---|---|---|---|
| GLM 5.3 Xploded glm-5.3-xploded | GLM 5.3 uncensored API | 1M | $2.00 / $5.00 | tools, reasoning |
| GLM 5.3 Flash Xploded glm-5.3-flash-xploded | GLM 5.3 Flash uncensored API | 1M | $0.30 / $1.00 | tools, vision, reasoning |
| DeepSeek V4.1 Flash Xploded deepseek-v4.1-flash-xploded | DeepSeek V4.1 Flash uncensored API | 1M | $0.30 / $1.50 | tools, reasoning |
| MiMo V2.6 Flash Xploded mimo-v2.6-flash-xploded | MiMo V2.6 Flash abliterated API | 1M | $1.00 / $2.00 | tools, vision, reasoning |
| Qwen3.8 27B Xploded qwen3.8-27b-xploded | Qwen3.8 27B uncensored API | 1M | $1.50 / $5.00 | tools, vision, reasoning |
| Qwen3.8 27B Fast Xploded qwen3.8-27b-fast-xploded | Qwen3.8 27B Fast uncensored API | 524K | $1.50 / $5.00 | tools, vision, reasoning |
FAQ
Does Fireworks AI allow NSFW or uncensored content?
Its terms say input must not include nudity or other sexually suggestive content, or content they consider harmful or objectionable at their discretion, and the serverless catalogue carries no uncensored or abliterated models. MagmaRouter's Xploded line is uncensored by design.
Can I use Claude Code with Fireworks AI?
Yes. Fireworks documents an Anthropic-compatible /v1/messages endpoint for Claude Code, though count_tokens and server-side tools are not supported. MagmaRouter also offers an Anthropic-compatible /v1/messages endpoint.
Which is better for fine-tuning or hosting my own model?
Fireworks, clearly. It supports uploading custom models and LoRAs on dedicated GPUs and runs SFT and DPO fine-tuning. MagmaRouter only sells its fixed Xploded line and does not host your models.
Which uncensored models does MagmaRouter sell?
MagmaRouter sells GLM 5.3 Xploded, GLM 5.3 Flash Xploded, DeepSeek V4.1 Flash Xploded, MiMo V2.6 Flash Xploded, Qwen3.8 27B Xploded, Qwen3.8 27B Fast Xploded. The table above has the price and context for each.
Are these models abliterated or uncensored?
MiMo V2.6 Flash Xploded is abliterated: the refusal direction was removed from the weights. GLM 5.3 Xploded, GLM 5.3 Flash Xploded, DeepSeek V4.1 Flash Xploded, Qwen3.8 27B Xploded, Qwen3.8 27B Fast Xploded are uncensored: the refusal was trained back out. The abliterated vs uncensored page explains the difference.
Which one answers fastest?
MiMo V2.6 Flash Xploded usually starts answering in one to two seconds. The GLM 5.3 Flash model thinks before it answers, so its first word can take much longer.