We're happy to announce that we've enlisted Fireworks AI as PayPerQ's primary provider for open-source models.

In AI inference, quality, latency, reliability, and privacy vary a lot — even when two providers are serving the exact same open weights. The model file is identical; the engine running it, the hardware it runs on, and the way requests are batched and routed are not. Those differences show up as slower first tokens, inconsistent output quality, timeouts under load, and murky data-handling policies. After putting providers through their paces, Fireworks earned our trust across all four dimensions, and we're confident bringing their stack to our users.

What Is Fireworks AI?

Fireworks AI is a production inference platform purpose-built for serving open-source models at scale. Rather than being a lab that trains its own frontier model, Fireworks specializes in one thing: running the best open-weight models — Llama, Qwen, DeepSeek, Kimi, GLM, Mistral, and more — as fast, reliably, and cost-effectively as possible.

The platform processes 30+ trillion tokens per day across 100+ hosted models, and it's the production inference layer behind well-known AI products. Cursor chose Fireworks because it's "way more performant than the open source engines." Notion cut latency from roughly 2 seconds to 350 milliseconds. Quora reported a 3x speedup in response time after migrating. That kind of engineering — custom serving stacks, optimized attention kernels, and elastic global scaling — is exactly what separates a great inference provider from a merely adequate one.

Why We Chose Fireworks

The tweet announcing this partnership named four things that "vary a lot" in AI inference. Here's how Fireworks delivers on each:

• Quality — Serving open weights correctly is harder than it looks. Quantization choices, tokenizer handling, and sampling defaults can quietly degrade output. Fireworks runs models at production-grade fidelity so the responses you get match the model's true capability.
• Latency — Fast time-to-first-token and high sustained throughput mean snappier chats and quicker agentic loops, not just better benchmark numbers.
• Reliability — Elastic, global scaling keeps models responsive under load. Fewer timeouts, fewer failed requests, more consistent behavior when it matters.
• Privacy — Fireworks offers enterprise-grade data controls and doesn't train on customer traffic — aligning with PayPerQ's privacy-first philosophy, where conversations are stored locally by default and accounts are optional.

Major Models Fireworks Serves on PayPerQ

Fireworks hosts a deep catalog of open-source models. Here are the major families now served through PayPerQ:

Model Family	Notable Variants	Best For
Llama	Llama 3.3 70B, Llama 3.2, Llama 3.1, Llama Vision	Fast general-purpose chat, multilingual, vision
Qwen	Qwen3 (0.6B–480B), Qwen3-VL 235B, Qwen Coder, QwQ 32B	Coding, reasoning, vision-language, long context
DeepSeek	DeepSeek V4-Pro, V3.2, R1, Prover V2	Reasoning, math/proofs, cost-efficient frontier quality
Kimi	Kimi K2.7 Code, K2.6, K2 Thinking	Agentic coding, long-horizon tool use, multimodal
GLM	GLM 5.2, GLM 5.1, GLM 4.7	Agentic engineering, tool calling, long context
MiniMax	MiniMax M3, M2.7	Long-context reasoning and generation
Mistral / Mixtral	Mistral Large 3, Ministral 3, Mixtral 8x22B MoE	Efficient European-built general models
gpt-oss	gpt-oss-120B	Open reasoning and function calling
Gemma	Gemma 4, Gemma 3	Compact vision-language, many-language support

Context lengths across these models range from a few thousand tokens up to 1M tokens on the largest variants, and the lineup spans everything from lightweight fast chat to heavyweight reasoning, coding, and vision workloads. As Fireworks adds new open-weight releases — and the open-source frontier moves fast — you'll see them appear on PayPerQ.

Key Capabilities

• Access the best open-source models in one place, served on infrastructure tuned for speed and reliability
• Coding and agentic work with Kimi, Qwen Coder, GLM, and DeepSeek — strong tool use and long-horizon reasoning
• Reasoning and math with DeepSeek R1, QwQ, and gpt-oss-120B
• Vision-language tasks with Qwen3-VL, Llama Vision, and Gemma 4
• Long-context work — summarize, analyze, and reason over large documents on high-context variants

And because PayPerQ is pay-per-query, you can switch between all of these models freely and only pay for what you actually use.

What is PayPerQ?

PayPerQ is the private way to use AI. You get the big-name models from Anthropic, OpenAI, Google and xAI, the best open-source models, and TEE models that run inside a hardware enclave where not even the model's host can see your data. Chat, image, video and audio, all in one place, with no account and no subscription. You top up a balance with crypto or a card, from as little as 10 cents, and pay only for what you use. Most users spend about $4 a month, against $20+ for a single ChatGPT subscription.

Privacy is built in rather than bolted on. Every chat request is served inside a secure enclave that PayPerQ's servers, staff and logs cannot read into, so PayPerQ never sees your prompts or responses. Choose a TEE model and the model's host cannot read them either.

Why Use PayPerQ?

• Private by default - chats are served inside a secure enclave PayPerQ cannot read, and every chat model shows its privacy tier (Anon, ZDR or TEE)
• No account needed - no sign-up and no email; a credit id is all it takes
• Pay with crypto or card - Bitcoin, Lightning, Monero, stablecoins and more, or a credit card
• Hundreds of AI models from every major provider in one place
• Pay per use - no subscriptions; top up from 10 cents and pay only for what you use, about 1 cent per query on average
• Your history stays yours - conversations are stored on your device by default

Getting Started

1. Visit ppq.ai
2. Top up your balance with crypto or a card (as little as 10 cents)
3. Pick a model, check its privacy tier, and start chatting

No account, no sign-up. Fund and go.