Korai serves open-weight models chosen on public benchmarks. No proprietary model (GPT, Claude, Gemini) ever runs on our network: only open, auditable models with publicly downloadable weights.
Every 30 days, we audit the fleet against public benchmarks (ArtificialAnalysis Intelligence Index, Open LLM Leaderboard, Chatbot Arena). If a new open-weight model beats a current fleet member, it gets added or swapped in. The rotation is automatic across every GPU on the network. No action needed on your end.
Open-weight, commercial license (Apache 2.0, MIT)
Top-3 in its category on public benchmarks
Runs on accessible hardware (consumer GPU or via an EU API)
Today, the core of Korai's fleet runs on the Qwen 3.8 family (fast, balanced, code, reasoning) and Gemma 4 26B A4B (multimodal, efficiency). Kimi K3, the largest open-weight model ever published, rounds out the offering in Ultra Deep mode for the most demanding use cases.
Alibaba · Apache 2.0
The workhorse of Korai's fleet: general use, code, and fast responses. Impressive numbers on Alibaba's own benchmark table, pending independent reproduction on our own evaluations.
27 billion parameters (Qwen 3.8 family)
Not publicly disclosed yet
Deployment being qualified on the Korai network
Alibaba · Apache 2.0
The reasoning variant of Qwen 3.8-27B: longer chain-of-thought, holds up better on multi-step tasks (terminal/tool use, complex instruction-following).
27 billion parameters, reasoning variant of the Qwen 3.8 family
Not publicly disclosed yet
Deployment being qualified on the Korai network
Google DeepMind · Apache 2.0
An efficient MoE variant of Gemma 4: lower compute cost (4B active) for intelligence close to heavier dense models. Inherits the Gemma 4 family's native multimodality (text, image, OCR).
Mixture-of-Experts · 26 billion total parameters · 4 billion active per token
Up to 256K tokens (inherited from the Gemma 4 family)
Deployment being qualified on the Korai network
Moonshot AI · Modified MIT (commercial use permitted)
The largest open-weight model ever published (2.8 trillion total parameters), released July 27, 2026. Kimi Delta attention makes long-context inference up to 6x cheaper by its own published claim. Multimodal, modified MIT license permitting commercial use.
MoE · 2.8 trillion total parameters · 104 billion active · Kimi Delta attention
1,000,000 tokens
Reserved for Ultra Deep mode, served by the network's most powerful nodes. Available to Pro and Max subscribers.
A comparison on the two benchmarks shared by both families, plus each model's own scores.
| Model | GPQA Diamond | LiveCodeBench v6 | Other benchmarks |
|---|---|---|---|
| Qwen 3.8-27B | 89.2 | 90.3 | SWE-bench Pro 61.7 · Terminal-Bench 2.1 73.0 · IFBench 79.5 |
| Gemma 4 26B A4B | 82.3 | 77.1 | MMLU Pro 82.6 · AIME 2026 88.3 · Codeforces Elo 1718 |
| Kimi K3 | — | — | Artificial Analysis Intelligence Index 57 · #1 Frontend Code Arena (Elo 1679) |
Qwen 3.8-27B: Alibaba's own benchmark table, not yet independently reproduced. Gemma 4 26B A4B: Google model card, April 2026. Kimi K3: Artificial Analysis (Intelligence Index) and the Frontend Code Arena (public leaderboard), model released July 27, 2026. Vendor figures, public benchmarks; updated on every fleet rotation.
When Korai adds, replaces, or retires a model, the change rolls out automatically across every GPU on the network, with no human intervention and no service interruption for you.
Korai pushes a new Ed25519-signed manifest. No outside machine can substitute a malicious model.
Every GPU on the network polls the manifest every 5 minutes. The signature is checked against a locally pinned key.
The new model is pulled from HuggingFace while the old one keeps serving your prompts. Zero downtime.
Once the download completes, the host swaps its subprocess. The orchestrator routes elsewhere during the 30-second transition.
Old models stay in disk cache for 24h (rollback possible if an issue is detected), then get automatically purged.
Korai tracks these models for addition to the fleet as soon as they ship or hardware conditions allow it.
Open the chat and switch between modes. Automatic routing picks the best model for your prompt.