Laptop GPU
6 GB
Runs today with Q5_K_M, a 24K window and an 8-bit KV cache.
34 tok/s
measured on an RTX 4050 Laptop GPU
Your AI.Your hardware.Your control.
KOLDOS is a programming assistant that runs on your own computer. Your GPU runs the model, and the tools, the memory and every trace stay on your disk. There is no cloud mode.
This computer
Reading hardwareReading- Operating system
- ···
- CPU threads
- ···
- Memory
- ···
- GPU
- ···
- VRAM
- ···
Estimate at 4-bit, 4K context
Waiting for the browser to report hardware.
Digital brain
How a KOLDOS turn thinks
Every request becomes a turn. KOLDOS decides a route, gathers only the context it needs, runs the model on your GPU, acts through tools and checks the result before it answers.
How the system fits togetherRoute
A multi-step, multi-file goal. Five to forty model calls, bounded by the turn's budget.
Stages on this route
RequestIn code
From the desktop app, the CLI or the local API.
Drawn from the real KOLDOS architecture. It shows how a turn moves, not live activity. On your machine, koldos brain draws the same graph from real events only.
- tok/s measured on a 6 GB laptop GPU
- 34
- token context window
- 24,576
- tools behind one boundary
- 25
- active context target, in tokens
- <10K
Hardware
Built for a mid-range GPU
KOLDOS is designed around an 8 GB graphics card and 24 GB of RAM, and already runs on a 6 GB laptop GPU. It picks a runtime profile for the memory it finds.
Check your exact configurationDesign target
8 GB
The mid-range machine KOLDOS is designed around, with 24 GB of RAM.
47.5 tok/s
Q5_K_M at 4K depth on an RTX 5050
Larger profile
10 GB+
The runtime profile moves to Q6_K with a 32K window.
Q6_K · 32K
runtime profile, not a benchmark
Minimum VRAM for each model at 4-bit
Context
7B · 4-bit · 5.2 GB
22B · 4-bit · 14.4 GB
7B needs system RAM
7B fits on the GPU
7B and 22B fit
6 GB
RTX 4050 Laptop
8 GB
RTX 4060
RTX 5060
RTX 4070 Laptop
12 GB
RTX 3060 12 GB
RTX 4070
RTX 5070
16 GB
RTX 4080
RTX 5080
RX 7800 XT
24 GB
RTX 4090
RX 7900 XTX
32 GB
RTX 5090
Estimates from model size, quantization and context length. Measured runs are in the benchmark database. How the estimate works.
Models
Different sizes, different jobs
KOLDOS 7B is the model KOLDOS runs today. A 22B model is only a future plan, with no release scheduled. More parameters is not automatically better.
- KOLDOS 7B
- Available
The model KOLDOS runs today.
- KOLDOS 22B
- Future
Possible future generation, in planning.
Size of the weights, drawn to scale
KOLDOS 7B
Available4.0 GB
KOLDOS 22B
Future12.4 GB
Weights only, at about 4.85 bits per weight (Q4_K_M-style). Context and runtime need more on top.
Capabilities
What works today, and what doesn't yet
Every capability carries one of five states. The same labels appear on every page, so nothing is described as working before it does.
All capabilities in detailWorks today9
- ChatAvailableTalk to KOLDOS in a conversation. Every message is a turn answered by the model on your machine.
- CodingAvailableReads, writes and edits files in your project, runs commands and checks the result.
- Local executionAvailableThe model, the tools and the memory all run on your computer. There is no cloud mode.
- ToolsAvailable25 tools inside a bounded loop: files, project search and shell commands, exposed only when a turn needs them.
- FilesAvailableAttach files to a turn and browse the project tree.
- Local APIAvailableA REST API with streaming on your own machine, giving any compatible client the whole KOLDOS turn.
- MemoryExperimentalProject rules, preferences and past corrections in a local database, isolated per project.
- VisionExperimentalReads screenshots, diagrams and charts through OCR and a small vision model on the CPU.
- Web searchExperimentalCurrent information brought into the turn as evidence marked untrusted.
Not yet3
- Available
- Works in the current build.
- Experimental
- Usable today, but behavior and interfaces can still change.
- In development
- Being built now. Not usable yet.
- Planned
- Designed or committed to, but work has not started.
- Future
- Long-term or on hold. Not committed.
Architecture
Where each part runs
Local by design. Only web search reaches the internet during a turn, and the hosted API is still a plan.
How the system fits togetherOn your computer
Model, tools, memory and traces
- Kernel
- Decision Engine
- Context Broker
- Inference
- Executor
- Verifier
- Chat
- Coding
- Local execution
- Tools
- Memory
- Vision
- Files
- Local API
Reaches outside
Needs the internet
- Web search
Planned
Designed, not built
- Hosted API
- Automation
- Agents
Developers
A whole turn behind one request
- Local API
- Available
- Hosted API
- Planned
The local API gives any client that speaks the common chat completions format the complete KOLDOS turn: routing, tools and verification, not just the model. The hosted API is planned with the same format.
Pricing
KOLDOS API plans
Four plans for the hosted KOLDOS API, paid with credits. They don't apply to KOLDOS Local, which runs on your own hardware without credits or subscriptions. Access is granted by hand.
Pricing structure- FreeP1Trying KOLDOS as a service.Queue priority$0100 credits per week
- PremiumP5Regular use with a higher place in the queue.Queue priority$4.99/month500 credits per day
- DeveloperP6Building on KOLDOS from your own code.Queue priority$9.99/month1,500 credits per day
- APIP7Applications and integrations that call KOLDOS with API keys, including Discord bots.Queue priority$14.99/month3,000 credits per day
Roadmap
Where KOLDOS is going
No dates, because none have been committed. Every column is driven by the same status data as the rest of the site.
Full roadmapNow10 items
- KOLDOS 7B
- Chat
- Coding
- Local execution
- +6
Next3 items
- Long-term memory
- Agent loop
- Evaluation series
Planned8 items
- Hosted API
- Automation
- Agents
- Auxiliary models
- +4
Research4 items
- KOLDOS 22B
- Model specialization
- 22B-class models on 8 GB GPUs
- Modular brain