Which Fikra API model is right for my workload?
Fikra API abstracts the complexity of hardware provisioning by routing your requests across pinned third-party open-weight and proprietary models seamlessly. The registry allows you to explicitly declare your required pipeline based on latency, context window, and reasoning depth.
Model Family Overview
The Fikra model lineup is built for every stop on the price-performance curve: DeepSeek's
newest flash builds for everyday chat and agents, GLM and Kimi for heavy reasoning, MiMo for
budget vision work. To select a model, inject the exact
String ID into the
model
parameter of your request body. This page is generated server-side from the same registry
that powers GET /v1/models.
| Model String ID | Context Window | Input / 1M tok | Output / 1M tok | What it is |
|---|---|---|---|---|
| deepseek-v4.1-flash | 1,048,576 | $ 0.05 | $ 0.15 | |
| deepseek-v4.1-flash-1017 | 1,048,576 | $ 0.05 | $ 0.15 | The latest DeepSeek V4.1 Flash update (Oct 17 build): better reasoning, coding, agentic and vision scores, fas |
| deepseek-v4-flash-0731 | 1,048,576 | $ 0.05625 | $ 0.17 | |
| glm-5.3-flash | 327,680 | $ 0.155 | $ 0.49 | |
| kimi-k3 | 1,113,088 | $ 2.1 | $ 10.5 | |
| step-5-preview | 1,048,576 | $ 1.25 | $ 3.3 | |
| mimo-v2.6-pro | 1,048,576 | $ 0.5 | $ 1.0 | Xiaomi MiMo-V2.6-Pro: the top-scoring open-weight model on the Artificial Analysis index right now. Vision + r |
| deepseek-v4-flash-ark | 1,048,576 | $ 0.55 | $ 1.65 | DeepSeek V4 Flash served through BytePlus ModelArk. 1M context, 393k output. |
| gemini-4-argon | 1,048,576 | $ 2.5 | $ 12.0 | |
| glm-5.3 | 327,680 | $ 1.05 | $ 3.35 | Z.ai flagship GLM-5.3: large-scale reasoning for complex software engineering and long-horizon agent tasks. St |
| glm-5.2 | 262,144 | $ 0.75 | $ 1.65 | Z.ai GLM-5.2 on a 1M context: strong coding and reasoning at a lower price than GLM-5.3. The budget choice in |
| nvidia-nemotron-3.5-super-vl | 1,048,576 | $ 0.5 | $ 1.25 | |
| qwen3.8-27b | 1,001,536 | $ 0.045 | $ 0.35 | Fast 27B-class multimodal model with automatic prompt caching. Cached input never costs more than fresh input. |
| deepseek-v4-pro-ark | 1,048,576 | $ 1.65 | $ 4.95 | DeepSeek V4 Pro served through BytePlus ModelArk. 1M context, 393k output. |
| deepseek-v4.1-flash-ark | 1,048,576 | $ 0.1875 | $ 0.75 | DeepSeek V4.1 Flash served through BytePlus ModelArk. 1M context, 393k output. |
| deepseek-v4-1-flash-economy-free | 1,048,576 | $ 0.0 | $ 0.0 | Temporary free promotional route backed by BytePlus ModelArk. Available for two weeks or until the shared $400 |
| fikra-pro-20b | 131,072 | $ 0.2 | $ 0.2 | |
| fikra-pro-120b | 131,072 | $ 0.3 | $ 0.3 | |
| fikra-qwen-27b | 262,144 | $ 0.045 | $ 0.35 | |
| fikra-qwen-max | 1,048,576 | $ 6.0 | $ 12.0 | |
| fikra-auto | 262,144 | $ 0.045 | $ 0.35 | |
| fikra-auto-pro | 1,048,576 | $ 6.0 | $ 12.0 | |
| fikra-code-35b | 262,144 | $ 0.045 | $ 0.33 |
Choosing Between Long Contexts and Reasoning Models
Several models in the matrix (the DeepSeek and GLM lines, MiMo-V2.6-Pro) are reasoning
models: before the visible answer they run internal chain-of-thought. Reasoning tokens
arrive in usage.completion_tokens_details
and are billed the same as normal output tokens.
Reasoning models can spend a large number of hidden tokens before output starts. Always allocate generous max_tokens so the visible answer is not cut mid-thought, and raise your HTTP client timeout: time-to-first-token is intentionally slower on deep reasoning requests.
Need the full 1,000,000-token context windows? point at
deepseek-v4.1-flash or
deepseek-v4-flash-0731.
Just need cheap, fast chat? Start with
deepseek-v4.1-flash or
glm-5.3-flash.
Each model's own page links from the catalog
with per-model playgrounds.
Upstream Anonymity & Resilience
Fikra API automatically abstracts engine saturation. When a serving node is rate-limited, slow or erroring, the router walks internal high-availability alternatives to keep your connection alive. Fikra scrubs internal routing metadata before returning the payload, ensuring no backend details leak while maintaining high uptime.