Repository navigation
Releases
Releases · QuixiAI/QuixiGate
QuixiGate v0.2.1
Highlights in v0.2.1
8.4 msCold Start &23.7 MiBIdle RSS (14 MBphysical footprint):- Deduplicated all outbound TLS root certificate stores into a single shared
OnceLock<rustls::ClientConfig>(ring+webpki-roots) and converted allreqwest::Clientpools to lazy per-coreOnceLockinitialization (0HTTP pools allocated at boot). - Constructed the 349-route
axum::Routerservice tree once instead ofNtimes across per-core workers, right-sized the outer CLI runtime, and moved initial local-runtime port discovery off the startup critical path. - Lazy-loaded the embedded
models.dev.json.gzcatalog snapshot on first read.
- Deduplicated all outbound TLS root certificate stores into a single shared
- Reproducible Scale Benchmark Suite (
litellm_vs_quixigate_bench):- Added a 3-process apples-to-apples scale benchmark harness (
crates/quixigate-proxy/examples/litellm_vs_quixigate_bench.rs) verifying70,690 req/sstateless JSON,66,605 req/sgoverned virtual-key + budget JSON,3.09M tok/sgoverned SSE streaming, and974,742 tok/sacross1,000concurrent GPU streams in156.9 MiBpeak RSS with0errors.
- Added a 3-process apples-to-apples scale benchmark harness (
- HTTP/1.1 Keep-Alive Hygiene & Production Installers:
- Added
Connection: closeon early413 Payload Too Largerejections per RFC 9112 §9.6. - Native OS installers (
install.sh->/usr/local/bin/quixigateon macOS & Linux;install.ps1->%LOCALAPPDATA%\Programs\QuixiGate\bin\quixigate.exeon Windows) with SHA-256 verification.
- Added
Install / Upgrade
macOS & Linux:
curl -fsSL https://raw.githubusercontent.com/QuixiAI/QuixiGate/main/install.sh | shWindows (PowerShell):
irm https://raw.githubusercontent.com/QuixiAI/QuixiGate/main/install.ps1 | iexFull Changelog: v0.2.0...v0.2.1
v0.2.0
Quick Install
macOS & Linux:
curl -fsSL https://raw.githubusercontent.com/QuixiAI/QuixiGate/main/install.sh | shWindows (PowerShell):
irm https://raw.githubusercontent.com/QuixiAI/QuixiGate/main/install.ps1 | iexHighlights
Actix-Style Share-Nothing Per-Core Engine & Polyglot Zero-DOM Cut-Through
67,751 req/s(100%parity with a bare zero-copy Rust TCP pipe,1.76 msp99) and9.47M tok/s(1.56 GiB/s/13.4 Gbps) SSE streaming throughput on an 18-core machine.- Lock-Free In-Memory Data Plane + Batched SQLite WAL Flusher:
ArcSwap<ControlSnapshot>+ SHA-256[u8; 32]key indexing +AtomicI64NanoUSD reservations + linearizableSAVEPOINTWAL write-behind batching (0SQLite mutex locks on the inference hot path). - 3 Native Zero-DOM Cut-Through Lanes: Borrowed envelope peeks (
serde::de::IgnoredAny) + SIMDmemchrBytessplicers +CutThroughSseStream(0tokio::spawn,0mpsc,0per-token JSON DOM) for:POST /v1/chat/completions-> OpenAI / OpenRouter / Copilot / Qwen / local runtimesPOST /v1/responses-> ChatGPT Codex (/responses)POST /v1/messages-> Anthropic (anthropic/anthropic-token) & MiniMax (minimax-portal)
- Zero-Spawn Polyglot Transcode State Machines: Direct
BytesMutStreamdrivers across all 5 codecs (openai_chat,openai_responses,anthropic,gemini,vercel_ui),O(1)generation-stampedCachedModelPlan, concrete zero-box Tower middleware,64 KiBvectoredwritevHyper connection builder, LinuxSO_REUSEPORTmulti-queue listeners,mimalloc, and full Fat LTO (lto = "fat").
100% LiteLLM Endpoint Compatibility (997 / 997 Routes)
- Drop-in coverage across all
997LiteLLM route decorators: native inference, embeddings, models, spend/budget management, virtual keys, teams, users, organizations, guardrails, MCP, and health endpoints—or structured400responses with actionable zero-paywall DIY migration guidance.
Native Windows (x86_64-pc-windows-msvc) Support
- First-class native Windows
.exerelease alongside macOS (aarch64-apple-darwin) and Linux (x86_64-unknown-linux-gnu), including seamless.exe/.cmd/.batagent shim resolution inquixigate run.
Prebuilt Archives
- macOS (Apple Silicon):
quixigate-aarch64-apple-darwin.tar.gz - Linux (
x86_64):quixigate-x86_64-unknown-linux-gnu.tar.gz - Windows (
x86_64MSVC):quixigate-x86_64-pc-windows-msvc.zip
Full Changelog: v0.1.0...v0.2.0
v0.1.0
Full Changelog: https://github.com/QuixiAI/QuixiGate/commits/v0.1.0