Pick Fireworks AI for production text and vision LLMs; pick Replicate for multimodal and experimental models. The two host barely overlapping catalogs, so the choice is usually decided by what you are serving, not by price. Fireworks is a per-token LLM serving platform built for high-throughput text generation, with custom FireAttention kernels, per-token fine-tuning, and weight export. Replicate is a run-any-model-as-a-container platform whose strength is image, video, and audio, with one-line deploys via Cog, billed per GPU-second with real cold starts. Replicate is now Cloudflare-owned and folding into Workers AI.
The split decides the choice. Fireworks prices text by the token at $0.90 per million for a 70B model (as of early 2026), with batch and cached-input discounts, on a warm serverless pool. Replicate prices by the second of GPU time, from $0.81/hr on a T4 to $5.49/hr on an H100, and a cold start can burn 10 to 60 seconds of that compute before your request runs, but it serves FLUX, Stable Diffusion, and 50,000 other models Fireworks does not.
TL;DR
- Pick Fireworks AI if you are serving production text or vision LLMs. $0.90/M for 70B text models, FireAttention with FP8 and speculative decoding, 50% off batch and cached input, per-token fine-tuning with weight export, plus SOC 2 Type II and HIPAA.
- Pick Replicate if you need image, video, audio, or experimental models, or one-line Cog deploys of your own model. 50,000+ models behind one API, billed per GPU-second so cold starts and idle time count, now wiring into Cloudflare Workers AI.
Who Wins Per Workload
The catalogs barely overlap, so most decisions are clear-cut once you name the workload.
| Workload / decision | Fireworks AI | Replicate |
|---|---|---|
| Production text LLMs | Fireworks: per-token FireAttention | Replicate: not its focus |
| Image / video / audio | Fireworks: not hosted | Replicate: FLUX, SD, thousands more |
| Lowest latency floor (text) | Fireworks: warm pool, no cold start | Replicate: cold starts hit |
| Cheapest at sustained text volume | Fireworks: per-token, no idle cost | Replicate: idle GPU burns |
| Run your own model / engine | Fireworks: curated LLMs only | Replicate: any model via Cog |
| Fine-tune and export weights | Fireworks: per-token SFT/DPO | Replicate: trains, GPU-billed |
| Experimental / niche models | Fireworks: limited | Replicate: 50,000+ catalog |
| Strictest compliance | Fireworks: SOC 2, HIPAA, ISO | Replicate: Cloudflare footprint |
| Fastest first call | Fireworks: warm serverless | Replicate: 10-60s cold start |
Quick Comparison
Fireworks optimizes one thing well: serving open text and vision LLMs by the token, fast. Replicate optimizes breadth and modality: any model, packaged once, called via one API, with image and video as its center of gravity.
| Spec | Fireworks AI | Replicate | Morph |
|---|---|---|---|
| Primary Focus | Per-token text/vision LLM serving | Per-second run-any-model platform | Code apply (not a general host) |
| Billing Unit | Per 1M tokens | Per GPU-second | Per request |
| 70B Text Price | $0.90 / 1M tokens | Per GPU-second (varies) | N/A, not a general host |
| On-Demand H100 | $7.00/hr | $5.49/hr | N/A, not a general host |
| Model Catalog | ~100s of LLMs (text/vision) | 50,000+ (incl. image/video/audio) | Code-specific models only |
| Multimodal (image/video/audio) | No | Yes | No |
| Fine-Tune + Export Weights | Yes (per-token SFT/DPO) | Trains, GPU-billed | N/A, not a general host |
| Cold Starts | Low (warm serverless) | 10-60s on cold models | Managed fleet |
| Best For | Production text/vision LLMs | Multimodal + experimental models | Applying model-generated edits |
Billing Model: Tokens vs GPU-Seconds
This is the deepest difference between the two, and it decides most of your bill.
Fireworks bills per token. You send a request, you pay for input and output tokens, and you never see a GPU. Pricing scales by model size: $0.10/M for models under 4B, $0.20/M for 4B to 16B, $0.90/M for models over 16B, and separate MoE tiers ($0.50/M up to 56B, $1.20/M from 56B to 176B). Cached input is 50% off and batch jobs are 50% off both directions.
Replicate bills per GPU-second. Public community models charge only active processing time, but private deployments bill setup, idle, and active time. That means a dedicated H100 sitting idle waiting for occasional requests still bills at $5.49/hr. Replicate's Official Models (a curated subset) use predictable per-token or per-image pricing instead, which removes cold-start overhead for those specific models.
The Idle-Cost Trap
On Replicate, a private deployment running at 80%+ utilization costs only marginally more than a public model. But a deployment sitting idle burns the full GPU-hour rate whether or not it is serving traffic. For bursty or low-volume workloads, that idle time is where the bill grows. Fireworks serverless avoids this by billing tokens, not seconds.
Pricing: Fireworks Wins for Text, Replicate Wins for Breadth
For LLM text generation, Fireworks is cheaper and more predictable. For everything else (image, video, audio, custom models), Replicate is often the only option of the two.
| Metric | Fireworks AI | Replicate |
|---|---|---|
| Model under 4B | $0.10 / 1M tokens | Per GPU-second |
| Model over 16B (e.g. 70B) | $0.90 / 1M tokens | Per GPU-second |
| MoE 56B-176B | $1.20 / 1M tokens | Per GPU-second |
| Claude 3.7 Sonnet | Not hosted | $3.00 / $15.00 per 1M |
| DeepSeek R1 | Tier-priced | $3.75 / $10.00 per 1M |
| T4 GPU | Not exposed | $0.81/hr |
| A100 80GB | Not exposed | $5.04/hr |
| H100 (on-demand) | $7.00/hr | $5.49/hr |
| B200 (on-demand) | $10.00/hr | Not offered |
| Cached Input Discount | 50% | N/A |
| Batch Discount | 50% | N/A |
| FLUX schnell image | Not hosted | $0.003 / image |
Fireworks lists on-demand GPUs too, at $7.00/hr for H100/H200, $10.00/hr for B200, and $12.00/hr for B300. Those rates are higher than Replicate's GPU-hour pricing, but Fireworks' main product is the per-token serverless API, where you never rent a GPU at all.
Replicate has no monthly subscription and no free tier beyond a small "Try for Free" collection. New accounts use prepaid credit that expires after 12 months. A failed generation still costs GPU time.
Cost on a Real Workload
Cost on a real workload (computed from list prices, early 2026)
Serving a 70B text model at 50M output tokens/day. On Fireworks serverless that is 50 x $0.90 = $45/day = ~$1,350/mo, with no GPU to rent and no idle charge. Replicate has no per-token 70B SKU, so the equivalent is a dedicated H100 at $5.49/hr = ~$132/day = ~$3,953/mo per GPU running 24/7.
So the dedicated H100 only wins once it stays busy enough to push more than ~$3,953/mo of equivalent token volume, which at Fireworks' $0.90/M is about 4.4B output tokens/month, or ~1,700 output tok/s sustained per H100. Below that sustained rate the GPU sits partly idle and Fireworks serverless is cheaper. Above it, owning the hardware pays off. At 50M tokens/day (~580 tok/s averaged) the workload is well under break-even, so Fireworks wins this scenario by roughly 3x.
The arithmetic only applies to text. For image, video, or audio there is no Fireworks comparison to run, since it does not host those models.
Cold Starts: The Hidden Replicate Cost
Fireworks keeps serverless models warm, so cold starts are rarely your problem. On Replicate, cold starts are a line item.
| Behavior | Fireworks AI | Replicate |
|---|---|---|
| Serverless cold start | Low (warm pool) | Official models: 5-10s |
| Community model cold start | N/A | 10-30s |
| Custom/large model cold start | N/A | 60s+ |
| Billed during cold start? | No (token billing) | Public: no / Private: yes |
| Scale to zero | Serverless default | Private deployments bill idle |
On Replicate, public models do not bill setup time, but private deployments pay for the full boot sequence and idle time. A large custom model can take 60+ seconds to load into GPU memory before it serves a single token, and on a private deployment that loading time is on your invoice.
For latency-sensitive interactive traffic that calls a model many times per session, a 60-second cold start anywhere in the chain is fatal. Fireworks' warm serverless pool is the safer default; Replicate's cold starts are the price of running arbitrary containerized models on demand.
Neither is built for the coding-agent apply loop; if applying model-generated code edits is the bottleneck, that is a different tool (Morph Fast Apply, ~10,500 tok/s, with published benchmarks).
Model Catalog: 50,000 vs a Curated LLM Set
Replicate wins on breadth by a wide margin. Fireworks wins on depth for text.
Replicate hosts more than 50,000 models spanning language, image (FLUX, Stable Diffusion), video, audio, and upscaling. Its open-source Cog tool packages any model into a reproducible container, so the catalog grows from community contributions. After the Cloudflare acquisition, that catalog is being wired into Workers AI for edge deployment.
Fireworks focuses on serving popular open LLMs and vision-language models well: Llama, DeepSeek, Qwen, and others, each tuned on FireAttention. It is not a place to run a niche image upscaler or a one-off community model. It is a place to run a 70B chat model fast and cheap with an OpenAI-compatible API.
Inference Engine: FireAttention vs Cog Containers
Fireworks built an engine. Replicate built a packaging standard. Both are good at what they target.
FireAttention is Fireworks' proprietary serving stack: hand-written CUDA kernels, FP8 and FP4 quantization where quality allows, and adaptive speculative decoding. Fireworks reports 3 to 12x lower latency and up to 5.6x higher throughput than self-hosted vLLM on identical hardware. It also supports multi-LoRA adapter injection, structured JSON output, function calling, prompt caching, and "predicted outputs" for edit-and-rewrite workloads.
Replicate's core technology is Cog, the open-source tool that packages any model and its dependencies into a standard container with a uniform HTTP API. Cog does not optimize the model's kernels; it standardizes deployment. That is why Replicate can host 50,000 models, but also why raw token throughput on a given LLM is generally not its selling point.
Predicted Outputs
Fireworks' "predicted outputs" feature speeds up workloads where most of the output is already known, like rewriting a file with small edits. It is the closest thing either provider has to a code-edit optimization, but it is a general latency trick on a general model, not a code-specific apply endpoint.
Fine-Tuning: Fireworks Prices It, Replicate Hosts It
Fireworks publishes a clear fine-tuning price list. Replicate lets you train and host fine-tunes but bills the underlying GPU time.
| Model Size | LoRA SFT | Full SFT |
|---|---|---|
| Up to 16B | $0.50 | $1.00 |
| 16.1B-80B | $3.00 | $6.00 |
| 80B-300B | $6.00 | $12.00 |
| Over 300B | $10.00 | $20.00 |
Fireworks supports LoRA SFT, LoRA DPO, full SFT, and full DPO with per-token training pricing, then serves the result with multi-LoRA so you can host many adapters on one deployment. Replicate supports fine-tuning popular models (Llama, FLUX) and hosting the result, but billing reverts to GPU-seconds and, on private deployments, idle time.
Compliance & Enterprise
Fireworks carries the heavier compliance stack. Replicate inherits Cloudflare's enterprise footprint.
Fireworks is SOC 2 Type II and HIPAA compliant, and has achieved ISO 27001 (security), ISO 27701 (privacy), and ISO 42001 (AI management). Dedicated deployments run in logically isolated environments with no cross-customer access. For regulated workloads, that certification set is a real advantage.
Replicate, now part of Cloudflare, benefits from Cloudflare's global network and developer platform. The integration adds custom-model and fine-tune support to Workers AI and brings Replicate's catalog closer to the edge. If you already run on Cloudflare, that consolidation is a meaningful pull.
When to Use Fireworks AI
- Cheap, fast text generation. $0.90/M for 70B models, $0.10/M for small models, with 50% off batch and cached input. The lowest-friction way to serve open LLMs by the token.
- Latency-sensitive LLM traffic. FireAttention delivers 3 to 12x lower latency than self-hosted vLLM, with a warm serverless pool that avoids cold starts.
- Multi-LoRA deployments. Train many LoRA adapters at per-token prices, then serve them all on one deployment with adapter injection.
- Regulated workloads. SOC 2 Type II, HIPAA, and three ISO certifications (27001, 27701, 42001) cover security, privacy, and AI management.
- Structured output and tool use. Native JSON mode, function calling, and predicted outputs for edit-heavy workloads.
When to Use Replicate
- The widest model catalog. 50,000+ models across text, image, video, and audio, all behind one consistent API.
- Image and video generation. FLUX schnell at $0.003/image, FLUX 1.1 Pro at $0.04/image, plus thousands of community vision models Fireworks does not host.
- Packaging custom models with Cog. The open-source container standard makes deploying your own model a one-command push with a uniform HTTP API.
- Cloudflare-native stacks. Post-acquisition, Replicate's catalog and fine-tunes are integrating into Workers AI for edge deployment.
- Pay-per-second bursty jobs. For sporadic batch generation on public models, per-second billing with no monthly commitment fits the spend pattern.
Frequently Asked Questions
Is Fireworks AI or Replicate cheaper?
Cheaper depends on volume and modality. For steady text generation, Fireworks is cheaper because it bills per token: $0.90/M for a 70B model, with 50% off batch and cached input, and no cold-start charge. Replicate bills per GPU-second, so a cold start (10 to 60 seconds) plus idle time on a dedicated H100 at $5.49/hr can cost more than the equivalent token volume. For image, video, or audio generation that Fireworks does not serve, Replicate's pay-per-second model is the cheaper (and only) option of the two.
Did Cloudflare buy Replicate?
Yes. Cloudflare announced the acquisition in November 2025 and closed it in early 2026. Replicate stays a distinct brand, and its 50,000+ model catalog and Cog packaging tool are being integrated into Cloudflare Workers AI, adding custom models and fine-tunes to that platform.
What is FireAttention?
FireAttention is Fireworks' proprietary inference engine. It uses hand-written CUDA kernels, FP8 and FP4 quantization where quality allows, and adaptive speculative decoding. Fireworks reports 3 to 12x lower latency and up to 5.6x higher throughput than self-hosted vLLM on the same hardware.
Does Replicate support per-token LLM pricing?
Partially. Replicate's Official Models use per-token pricing (Claude 3.7 Sonnet at $3.00/$15.00 per million input/output, DeepSeek R1 at $3.75/$10.00). The rest of the 50,000-model catalog bills per GPU-second by hardware, so cold starts and idle time on private deployments count against you.
Can I export weights from Fireworks or Replicate?
Fireworks lets you fine-tune (LoRA SFT, LoRA DPO, full SFT, full DPO) at per-token training prices and serve the adapters, and it can export trained weights, so you are not locked in. Replicate's portability comes from Cog: any model is packaged as a standard container you can pull and run anywhere, the strongest portability story of the two for arbitrary models. Fine-tuning a text LLM and keeping the weights points to Fireworks; packaging and moving any model as a container points to Replicate's Cog.
Related Comparisons
Fireworks for Text LLMs, Replicate for Multimodal
Different catalogs, so the workload usually decides. If applying model-generated code edits is your bottleneck, that is a separate tool: Morph Fast Apply runs at ~10,500 tok/s with published benchmarks.