GPU prefill from day one
Krasis began in public with safetensors ingestion, cached INT4 weights and a GPU MoE prefill path. Fast prompt processing was part of the foundation, not a later add-on.
Public development record
Built in public.
Measured in public.
Krasis has evolved from an experimental GPU-prefill engine into a Rust/CUDA runtime spanning Ampere through Blackwell, one or many GPUs, native Windows, Linux and WSL2, and a growing set of MoE architectures.
09 Febfirst public commit
14launcher-supported models
Ampere+NVIDIA generations
3desktop platforms
February to August 2026
From an idea to a large-model runtime.
Dates follow the public repository. Every entry links to the commit or release that supports it; private experiments and work that has not shipped publicly are deliberately excluded.
Hybrid Qwen support and hot/cold residency
Qwen3-Coder-Next brought mixed Gated DeltaNet and attention layers. Two days later, HCS combined fast GPU prefill with a VRAM-resident hot expert set and on-demand access to the rest.
Decode becomes a GPU runtime
Resident attention, CUDA decode kernels and GPU MoE execution replaced the early CPU-first decode path. Per-layer CUDA graphs then removed repeated launch overhead from token generation.
N-GPU execution, Rust prefill and Ampere
Krasis added an arbitrary-GPU decode pipeline, moved prefill orchestration into Rust, and completed CUDA graph replay on Ampere—extending the production path beyond the newest NVIDIA generation.
Compact attention and KV caches
Native HQQ attention caches and compact KV formats reduced the non-expert memory floor. That created more room for resident experts while keeping quality-oriented and tighter-memory profiles distinct.
Krasis 1.0 ships
The first 1.0 release turned the experimental runtime into an installable product line, with managed setup, launcher-driven configuration and release-built packages.
The model frontier widens
Curated downloads and architecture work brought Gemma 4, Step-3.7-Flash and both 35B and 397B Ornith variants into the public runtime, alongside the growing Qwen and Nemotron families.
Vision and native Windows
Gemma 4 and Step vision paths landed alongside a native Windows installer. The Windows release uses a private, pinned runtime rather than modifying the user’s system Python.
GLM-5.2 sparse DSA validated
A new sparse attention and routing family joined Krasis, including native GPU routing and compact cache work. GLM-5.2 also pushed the supported model envelope to a much larger memory footprint.
DeepSeek-V4-Flash support lands
DeepSeek-V4-Flash-0731 added another distinct attention, cache and routing design. Subsequent public work optimized its prefill and decode paths from measured component timings.
Tools and sessions become first-class
Structured tool calls were normalized across model families, while exact prefix-state reuse and RAM-backed session snapshots made long, multi-turn workflows practical without replaying every prior token.
Topology becomes a measured runtime choice
Krasis added peer expert serving and then generalized launcher selection across available GPUs. It measures the topology and chooses between supported strategies instead of assuming every multi-GPU machine behaves alike.
Current launcher catalog
Different architectures.
One measured runtime.
Support means more than parsing a checkpoint. Each family has its own launcher profile, memory budget and validation boundary. The benchmark explorer shows only measured configurations and current evidence.
Qwen3-Coder-Next · Qwen3 235B · Qwen3.5 35B / 122B / 397B · Qwen3.6 35B
Nemotron-3 Nano 30B · Nemotron-3 Super 120B
Gemma 4 26B · text and vision
Step-3.7-Flash · text and vision
Ornith 1.0 35B · Ornith 1.0 397B
DeepSeek-V4-Flash
GLM-5.2