Understanding LLM Providers & Models in KForge
Last updated: 01/02/2026
This document explains how KForge thinks about AI providers, models, presets, and cost.
It is intentionally high-level and provider-agnostic.
Provider policies, pricing, and available models change frequently.
This document is a living reference and may be updated over time.
1. Providers vs Models (important distinction)
In KForge, providers and models are separate concepts.
- Provider: where the AI runs (cloud service, gateway, or local runtime)
- Model: the exact identifier that provider expects (case-sensitive)
KForge does not hardcode pricing, limits, or guarantees.
You are always in control of which providers and models you configure.
2. Types of providers
Cloud-native providers
These companies:
- build their own AI models
- run them on their own infrastructure
- enforce pricing and limits directly
Examples include OpenAI, Anthropic (Claude), Google (Gemini), and Mistral.
What this means:
- API key required
- Billing usually required
- Most stable and predictable behavior
- Model names and capabilities may still evolve
OpenAI-compatible gateways
These are independent services that:
- expose APIs compatible with the OpenAI format
- host their own models or route requests to multiple providers
Examples include OpenRouter, DeepSeek, and Groq.
Important:
“OpenAI-compatible” means same API shape, not same models, pricing, or guarantees.
Trade-offs:
- Often cheaper or temporarily free
- Free tiers may rotate or disappear
- Models may change without notice
- Excellent for experimentation and flexibility
Custom (bring-your-own endpoint)
The Custom provider allows you to connect:
- self-hosted models
- enterprise gateways
- proxies or internal services
Characteristics:
- Maximum control
- No discovery or marketplace
- KForge does not validate or manage pricing
This option is intended for advanced users who already know what they are connecting to.
Local runtimes
These run entirely on your own machine.
Examples include Ollama and LM Studio.
What this means:
- No API key
- No per-token billing
- Cost is hardware, electricity, and time
- Performance varies by system
Local models are ideal for:
- offline use
- privacy-sensitive workloads
- experimentation without cloud cost
3. About models and cost
There is no permanently “free” model.
Most cloud providers charge per token:
- input tokens (what you send)
- output tokens (what the model replies)
Even models labeled as “free” may:
- have rate limits
- expire
- require paid tiers later
For this reason, KForge treats all cost information as advisory only.
4. Model labels and usage modes in KForge
KForge may display model labels using color and usage mode.
These labels are guidance, not enforcement.
Cost (color labels)
- 🔵 Free — no billing required
- 🟢 Paid (low cost)
- 🟡 Paid (standard)
- 🔴 Paid (expensive)
- ⚪ Unknown — provider-dependent or unclear
Usage modes
- Sandbox — testing, quick iterations, throwaway work
- Main — day-to-day default usage
- Heavy — high capability; use sparingly
A model’s color indicates expected cost.
Its usage mode indicates when it should be used.
KForge never:
- modifies model IDs
- enforces pricing
- blocks usage based on cost
5. About presets
Presets are curated default models provided by KForge.
They exist to:
- reduce decision fatigue
- provide safe starting points
- demonstrate recommended usage patterns
Presets:
- are not exhaustive
- may overlap across providers
- can change over time
The current snapshot of presets is documented in:
Presets are defined in the app today but are expected to move to remote configuration in the future.
6. Duplicate models across providers (intentional)
You may see the same model name appear under multiple providers.
This is intentional.
Examples:
- the same model via a marketplace vs a direct provider
- the same model via a custom endpoint vs a hosted service
These options differ in:
- pricing
- availability
- routing
- privacy and control
The capability may be similar, but the tradeoffs are not.
7. Volatility and expectations
Some models are inherently volatile:
- preview releases
- rotating free tiers
- marketplace-provided “free” models
This is normal.
KForge is designed so that:
- models may appear or disappear
- users can adapt quickly
- documentation sets expectations clearly
Volatility is not a bug — it is part of the ecosystem.
8. Choosing a model (practical advice)
Instead of asking “what is the best model?”, consider:
- Am I experimenting or building something serious?
- Do I care more about speed or reasoning?
- Is this a temporary task or important logic?
You can switch models at any time.
There is no single correct choice.
9. Why this document lives outside the app
Provider models, pricing, and policies change constantly.
Embedding this information directly in the app would:
- become outdated
- mislead users
- require frequent rebuilds
By linking to documentation instead, KForge stays:
- honest
- lightweight
- future-proof
10. Core principle
KForge does not sell models.
KForge does not guarantee cost.
KForge gives you control.