RTX 3060 12GB
- Game Type
- Light / Competitive (Fortnite, Valorant, LoL)
- Experience
- ✅ Light voice mode; responsive co-hosting.Tested
OpenCohost Kira · Active development
OpenCohostKira
An AI co-host powered by PNG/VRM avatars, text-to-speech (TTS), memory, profiles, an agenda that lets the LLM iterate on itself, and Twitch stream chat ingestion.
Turn-based conversation
Requests are queued and answered in short turns instead of overlapping.
Voice in the runtime capture
Hear Kira's voice in the current product capture.
Failure recovery
See how the current interface recovers from an interaction failure.
Runtime video
Watch Kira in action
Uncut runtime video: turn-based voice and supervised recovery.
Audio: Spanish · English capture in progress
Preview upcoming v0.5.0-alphaIn Development: v0.5.0-alpha (Upcoming)Aug – Oct 2026 development
Preview the upcoming Kira v0.5.0-alpha
An advance look at the next major release: dedicated 3D VRM expressions, desktop companion overlay, integrated model management, and our local MiniLM v5 memory pipeline.
Star or follow the repository to track upcoming releasesStar on GitHub
Preview capture: desktop companion, lip-sync visemes, and in-app model downloader
Key v0.5.0 highlights
01
Dedicated VRM & Expression Engine
Custom 3D model designed specifically for Kira, featuring synchronized lip-sync and audio visemes.
02
Desktop Companion Mode
Interact with Kira in an independent, floating companion overlay synchronized with the main timeline.
03
In-App Model Downloader
Search and pull Ollama models directly from the application interface.
04
VRAM-Aware Context & Adaptive Reasoning
Automatic context scaling up to 16k tokens, spill protection, and adaptive reasoning budgets with manual ceilings.
05
Flexible Profile System
Clean, modular persona prompts removing legacy hardcoded constraints.
06
Memory v5
End-to-end all-MiniLM embedding integration for improved semantic indexing and recall. Local embedding — MiniLM-L12 ONNX served by an isolated background worker (no cloud, no extra VRAM on the LLM runner).
07
Speech Flow Controls
Option to discard audio residues on interruption and purge stale speech on fast regenerations.
08
/setup command
First-run LLM configuration (provider, Ollama, models) from the chat palette.
On screen
A supervised co-host for your live show.
Kira brings your avatar, chat context and prepared agenda together; you stay in control.
Editable profiles
Shape Kira with editable profiles.
Measured session
4 h 27 min of supervised agenda flow.
One instrumented run; not proof of production streaming or autonomous hosting.
- Exact duration 4:27:35
- 4 h 27 min
- agenda session
- Traced generations
- 285
- runtime total
- Cloud / local generations
- 120 / 165
- provider split
Local-first hardware matrix
Is your setup ready for Kira?
Local performance depends on your GPU, free VRAM, model, and game load.
RTX 3060 12GB
- Game Type
- Heavy / Ultra 3D (Cyberpunk, Alan Wake 2)
- Experience
- ⚠️ Light voice mode recommended; minor delay possible.Estimate
RTX 4090
- Game Type
- Any game / settings
- Experience
- ✅ More headroom for heavy modes; depends on model, game load, and features.Estimate
GTX 1660
- Game Type
- Lightweight games (Fortnite, Valorant)
- Experience
- ✅ Light voice; disable high-poly avatar.Estimate
No Dedicated GPU
- Game Type
- Just chatting screen / browsing
- Experience
- ❌ Local voice processing needs a dedicated GPU.Estimate
Only one PC is tested: RTX 3060 12GB. All other rows are estimates, not benchmarks; results depend on model, GPU, free VRAM, and game load.
Self-hosted alpha
Download the public alpha for Windows.
Windows alpha, published from CI. No hosted service or bundled model: bring Ollama, your chosen model, and suitable hardware.
v0.3.0-alpha.4 · 588 MB · Windows 10/11 x64 · published Sep 9, 2026 · Verify the SHA256 checksums
CI-built alpha. Report issues without API keys, private logs, or sensitive information.
Local alpha feedback
Help improve the alpha.
Share reproducible setup or workflow issues. Never include API keys, private logs, or sensitive information.
FAQ · 03
Three honest answers before you evaluate Kira
What is Kira today?
An alpha streaming co-host with supervised agenda, Direct/PTT, chat, memory and local/cloud model paths. It is not an autonomous host.
Can I use local or cloud models?
Both paths exist. Ollama has been tested with specific Gemma, Llama and Qwen configurations; BYOK supports cloud providers. Results depend on the exact model and hardware.
Is streaming production-ready?
No. It has not been personally validated in a real stream — only with external streamers by consuming their chats, and only as chat consumption, not as a proven pleasant experience for viewers. Twitch compatibility follows Twitch policies. YouTube is a different story: reading its chat uses a method YouTube Terms of Service do not allow, so using Kira there carries a real risk of a channel ban. In the app, live viewer chat is read from the Stream tab and remains in development: LLM responses can be inaccurate or inappropriate depending on your active audience profile, and whatever Kira says on your channel is your responsibility. If your LLM provider is cloud-based instead of local, viewer chat content leaves your machine.