Zumik
Sign inSign up

Fireworks AI

Open-weights serving with speculative decoding and dedicated tiers.

Fireworks AI caching guideBrowse all models
−74%
Cache-read discount
40%
Batch discount
237
Models on Zumik
Yes
BYOK supported

How caching works here

Fireworks serves open-weights models (DeepSeek, Llama, Kimi, Qwen, GLM, GPT-OSS) with prompt caching and speculative decoding. Dedicated deployments hold caches longer and give predictable latency, which is the bridge toward a full BYOC hot lane.

What Zumik sees

Fireworks is where managed serving meets BYOC. When a workload concentrates on one or two open-weights paths with strong locality, our replay runs frequently justify moving that lane to dedicated capacity or BYOC.

Pitfall

Running a hot, reuse-heavy lane on serverless and blaming the model for cold-cache latency, when a dedicated deployment would hold the prefix warm.

Profile
Min cache size512 tok
RetentionServerless idle window; dedicated holds longer
Service tiersserverless, dedicated
BYOCAvailable
open-weightshigh-volumeBYOC migration candidates

Fireworks AI models in the catalog.

ModelContextInputOutputCache readReuse-adj
Qwen3 4B128K$0.03$0.03$0.03 $0.03
Gemma 3 12B Instruct131K$0.05$0.10$0.05 $0.06
OpenAI gpt-oss-20b131K$0.07$0.30$0.04 −50%$0.11
OpenAI gpt-oss-safeguard-20b131K$0.07$0.30$0.04 −51%$0.12
Qwen3-VL-8B-Instruct131K$0.08$0.50$0.08 $0.18
MythoMax L2 13B4K$0.09$0.09$0.09 $0.09
Qwen3 235B A22B Instruct 2507131K$0.09$0.58$0.09 $0.21
Qwen3 30B-A3B41K$0.09$0.45$0.09 $0.18
Devstral-Small-2505128K$0.10$0.30$0.10 $0.15
Qwen3 30B A3B Instruct 2507128K$0.10$0.30$0.01 −90%$0.11
Gemma 3 27B Instruct98K$0.12$0.20$0.12 $0.14
DeepSeek-V4-Flash1M$0.14$0.28$0.03 −79%$0.13
DeepSeek R1 Distill Qwen 14B33K$0.15$0.15$0.15 $0.15
OpenAI gpt-oss-120b131K$0.15$0.60$0.01 −90%$0.21
Qwen3 8B131K$0.18$0.70$0.18 $0.31
Qwen3 VL 30B A3B Instruct131K$0.20$0.70$0.20 $0.33
Qwen3 VL 30B A3B Thinking131K$0.20$1.00$0.20 $0.40
Qwen3 Omni 30B A3B Instruct66K$0.25$0.97$0.25 $0.43
Deepseek V3 03-24164K$0.27$1.12$0.14 −50%$0.43
DeepSeek R1 Distill Qwen 32B64K$0.30$0.30$0.30 $0.30
MiniMax M2.7197K$0.30$1.20$0.06 −80%$0.43
Minimax M3512K$0.30$1.20$0.06 −80%$0.43
MiniMax-M2205K$0.30$1.20$0.03 −90%$0.41
NVIDIA Nemotron 3 Super 120B A12B BF16256K$0.30$0.90$0.30 $0.45
NVIDIA Nemotron 3 Super 120B A12B FP8256K$0.30$0.90$0.30 $0.45
Qwen3 235B A22B Thinking 2507131K$0.30$3.00$0.30 $0.97
Qwen3 VL 235B A22B Instruct131K$0.30$1.50$0.30 $0.60
Qwen3 14B131K$0.35$1.40$0.35 $0.61
Qwen3 Coder 30B A3B Instruct262K$0.45$2.25$0.45 $0.90
Deepseek R1 05/28164K$0.50$2.15$0.35 −30%$0.85
Kimi K2 Instruct131K$0.57$2.30$0.57 $1.00
Kimi K2 Instruct 0905131K$0.57$2.30$0.57 $1.00
Kimi K2 Thinking262K$0.60$2.50$0.15 −75%$0.89
Qwen3 235B A22B131K$0.70$2.80$0.70 $1.22
Qwen3 32B131K$0.70$2.80$0.70 $1.22
DeepSeek R1 Distill Llama 70B8K$0.80$0.80$0.80 $0.80
Kimi K2.6262K$0.95$4.00$0.16 −83%$1.39
Kimi K2.7 Code262K$0.95$4.00$0.19 −80%$1.40
Qwen3 VL 235B A22B Thinking131K$0.98$3.95$0.98 $1.72
GLM-5205K$1.00$3.20$0.20 −80%$1.22
DeepSeek V3131K$1.25$1.25$1.25 $1.25
GLM 5.1203K$1.40$4.40$0.26 −81%$1.68
Qwen3 Coder 480B A35B Instruct262K$1.50$7.50$1.50 $3.00
DeepSeek-V4-Pro1M$1.74$3.48$0.14 −92%$1.52
DeepSeek R1 (Fast)164K$3.00$7.00$3.00 $4.00
Chronos Hermes 13B v24K——— —
Code Llama 13B16K——— —
Code Llama 13B Instruct16K——— —
Code Llama 13B Python16K——— —
Code Llama 34B16K——— —
Code Llama 34B Instruct16K——— —
Code Llama 34B Python16K——— —
Code Llama 70B4K——— —
Code Llama 70B Instruct4K——— —
Code Llama 70B Python4K——— —
Code Llama 7B16K——— —
Code Llama 7B Instruct16K——— —
CodeGemma 2B8K——— —
CodeGemma 7B8K——— —
CodeQwen 1.5 7B66K——— —
Cogito v1 Preview Llama 3B131K——— —
Cogito v1 Preview Llama 70B131K——— —
Cogito v1 Preview Llama 8B131K——— —
Cogito v1 Preview Qwen 14B131K——— —
Cogito v1 Preview Qwen 32B131K——— —
DeepSeek Coder 1.3B Base16K——— —
DeepSeek Coder 33B Instruct16K——— —
DeepSeek Coder 7B Base4K——— —
DeepSeek Coder 7B Base v1.54K——— —
DeepSeek Coder 7B Instruct v1.54K——— —
DeepSeek Coder V2 Lite Base164K——— —
DeepSeek Coder V2 Lite Instruct164K——— —
DeepSeek Prover V2164K——— —
DeepSeek R1 (Basic)164K——— —
DeepSeek R1 0528 Distill Qwen3 8B131K——— —
DeepSeek R1 Distill Llama 8B131K——— —
DeepSeek R1 Distill Qwen 1.5B131K——— —
DeepSeek R1 Distill Qwen 7B131K——— —
DeepSeek V2 Lite Chat164K——— —
DeepSeek V2.533K——— —
DeepSeek V3.1164K——— —
DeepSeek V3.1 Terminus164K——— —
Deepseek v3.2164K——— —
Dolphin 2.6 Mixtral 8x7b33K——— —
Dolphin 2.9.2 Qwen2 72B131K——— —
ERNIE-4.5-21B-A3B-PT131K——— —
FARE-20B131K——— —
Firesearch OCR V68K——— —
Gemma 2 9B Instruct8K——— —
Gemma 2B Instruct8K——— —
Gemma 3 4B Instruct131K——— —
Gemma 4 31B IT NVFP4262K——— —
Gemma 4 E4B131K——— —
Gemma 7B8K——— —
Gemma 7B Instruct8K——— —
GLM-4.5131K——— —
GLM-4.5-Air131K——— —
GLM-4.5V131K——— —
GLM-4.6203K——— —
GLM-4.7203K——— —
GLM-4.7 Flash203K——— —
Hermes 2 Pro Mistral 7B33K——— —
InternVL3 38B16K——— —
InternVL3 78B16K——— —
InternVL3 8B16K——— —
KAT Dev 32B131K——— —
KAT Dev 72B Exp131K——— —
Kimi K2.5262K——— —
Llama 2 13B4K——— —
Llama 2 13B Chat4K——— —
Llama 2 70B4K——— —
Llama 2 7B4K——— —
Llama 2 7B Chat4K——— —
Llama 3 70B Instruct8K——— —
Llama 3 70B Instruct (HF version)8K——— —
Llama 3 8B8K——— —
Llama 3 8B Instruct8K——— —
Llama 3 8B Instruct (HF version)8K——— —
Llama 3.1 405B Instruct131K——— —
Llama 3.1 70B Instruct131K——— —
Llama 3.1 8B Instruct131K——— —
Llama 3.1 Nemotron 70B131K——— —
Llama 3.2 11B Vision Instruct131K——— —
Llama 3.2 1B131K——— —
Llama 3.2 1B Instruct131K——— —
Llama 3.2 3B131K——— —
Llama 3.2 3B Instruct131K——— —
Llama 3.2 90B Vision Instruct131K——— —
Llama 3.3 70B Instruct131K——— —
Llama 4 Maverick Instruct (Basic)1M——— —
Llama 4 Scout Instruct (Basic)1M——— —
Llama Guard 3 8B131K——— —
Llama Guard 7B4K——— —
Llama Guard v2 8B8K——— —
Llama Guard v3 1B131K——— —
MiniMax-M2.1197K——— —
MiniMax-M2.5197K——— —
Ministral 3 14B Instruct 2512256K——— —
Ministral 3 3B Instruct 2512256K——— —
Ministral 3 8B Instruct 2512256K——— —
MiroThinker-1.7262K——— —
Mistral 7B33K——— —
Mistral 7B Instruct v0.233K——— —
Mistral 7B Instruct v0.333K——— —
Mistral 7B OpenOrca33K——— —
Mistral 7B v0.233K——— —
Mistral Large 3 675B Instruct 2512256K——— —
Mistral Nemo Base 2407128K——— —
Mistral Nemo Instruct 2407128K——— —
Mistral Small 24B Instruct 250133K——— —
Mixtral 8x7B v0.133K——— —
Mixtral Moe 8x22B66K——— —
Mixtral MoE 8x22B Instruct66K——— —
Mixtral MoE 8x7B Instruct33K——— —
Mixtral MoE 8x7B Instruct (HF version)33K——— —
Molmo2-4B37K——— —
Molmo2-8B37K——— —
Nous Capybara 7B V1.933K——— —
Nous Hermes Llama2 13B4K——— —
Nous Hermes Llama2 70B4K——— —
Nous Hermes Llama2 7B4K——— —
Nouse Hermes 2 Mixtral 8x7B DPO33K——— —
NVIDIA Nemotron 3 Nano Omni 30B A3B262K——— —
NVIDIA Nemotron 3 Super 120B A12B NVFP4262K——— —
NVIDIA Nemotron 3 Ultra BF16262K——— —
NVIDIA Nemotron 3 Ultra NVFP4262K——— —
NVIDIA Nemotron Nano 12B v2128K——— —
NVIDIA Nemotron Nano 2 VL131K——— —
NVIDIA Nemotron Nano 9B v2128K——— —
OpenAI gpt-oss-safeguard-120b131K——— —
OpenChat 3.5 01068K——— —
OpenHermes 2 Mistral 7B33K——— —
OpenHermes 2.5 Mistral 7B33K——— —
Phi-3 Mini 128k Instruct131K——— —
Phi-3.5 Vision Instruct32K——— —
Phind CodeLlama 34B Python v116K——— —
Phind CodeLlama 34B v116K——— —
Phind CodeLlama 34B v216K——— —
Pythia 12B2K——— —
Qwen 3 4B Instruct 2507262K——— —
Qwen 3.5 122B A10B262K——— —
Qwen 3.5 35B A3B262K——— —
Qwen QWQ 32B Preview33K——— —
Qwen1.5 72B Chat33K——— —
Qwen2 72B Instruct33K——— —
Qwen2 7B Instruct33K——— —
Qwen2-VL 2B Instruct33K——— —
Qwen2-VL 72B Instruct33K——— —
Qwen2-VL 7B Instruct33K——— —
Qwen2.5 0.5B Instruct33K——— —
Qwen2.5 1.5B Instruct33K——— —
Qwen2.5 14B131K——— —
Qwen2.5 14B Instruct33K——— —
Qwen2.5 14B Instruct33K——— —
Qwen2.5 32B131K——— —
Qwen2.5 32B Instruct33K——— —
Qwen2.5 72B131K——— —
Qwen2.5 72B Instruct33K——— —
Qwen2.5 7B131K——— —
Qwen2.5 7B131K——— —
Qwen2.5 7B Instruct33K——— —
Qwen2.5-Coder 0.5B33K——— —
Qwen2.5-Coder 0.5B Instruct33K——— —
Qwen2.5-Coder 1.5B33K——— —
Qwen2.5-Coder 1.5B Instruct33K——— —
Qwen2.5-Coder 14B33K——— —
Qwen2.5-Coder 14B Instruct33K——— —
Qwen2.5-Coder 32B33K——— —
Qwen2.5-Coder 32B Instruct33K——— —
Qwen2.5-Coder 32B Instruct 128K131K——— —
Qwen2.5-Coder 32B Instruct 32K RoPE33K——— —
Qwen2.5-Coder 32B Instruct 64k66K——— —
Qwen2.5-Coder 3B33K——— —
Qwen2.5-Coder 3B Instruct33K——— —
Qwen2.5-Coder 7B33K——— —
Qwen2.5-Coder 7B Instruct33K——— —
Qwen2.5-Math 72B Instruct4K——— —
Qwen2.5-VL 32B Instruct128K——— —
Qwen2.5-VL 3B Instruct128K——— —
Qwen2.5-VL 72B Instruct128K——— —
Qwen2.5-VL 7B Instruct128K——— —
Qwen3 0.6B41K——— —
Qwen3 1.7B131K——— —
Qwen3 30B A3B Thinking 2507262K——— —
Qwen3 Coder 480B Instruct BF16262K——— —
Qwen3.5 27B262K——— —
Qwen3.5 397B A17B262K——— —
Qwen3.5 9B262K——— —
Qwen3.6 27B262K——— —
Qwen3.6-35B-A3B262K——— —
QWQ 32B131K——— —
Rolm OCR128K——— —
Seed OSS 36B Instruct524K——— —
Snorkel Mistral PairRM DPO33K——— —
Step-3.7-Flash-NVFP4262K——— —
Toppy M 7B33K——— —
Zephyr 7B Beta33K——— —

Fireworks AI, answered.

How does Fireworks AI prompt caching work?

Fireworks serves open-weights models (DeepSeek, Llama, Kimi, Qwen, GLM, GPT-OSS) with prompt caching and speculative decoding. Dedicated deployments hold caches longer and give predictable latency, which is the bridge toward a full BYOC hot lane.

What discount does Fireworks AI caching give?

Cache reads on Fireworks AI are about 74% cheaper than list input price.

Does Fireworks AI support BYOK on Zumik?

Yes. You can bring your own Fireworks AI key, and provider-native caching, batch, and service tiers stay active under your account.

What is the common Fireworks AI caching mistake?

Running a hot, reuse-heavy lane on serverless and blaming the model for cold-cache latency, when a dedicated deployment would hold the prefix warm.

Route Fireworks AI the smart way.

Capture Fireworks AI's 74% cache-read discount and batch tier automatically through Zumik.

Fireworks AI caching guideCompare providers