HFLabs Product Portfolio
Specialized low-latency engines and high-assurance compute substrates designed for autonomous agents, metaheuristic optimization, and direct hardware execution.
LevelZero.NET
Native Intel GPU acceleration and managed Level Zero runtime with pre-compiled SPIR-V kernels for .NET 10.
Key Capabilities
- ✓ Single DLL: Managed API, native C shim, and 418 pre-compiled kernel modules across 11 Intel GPU targets
- ✓ Unified Shared Memory (USM) zero-copy buffer allocations with typed Span<T>
- ✓ Precompiled modules for Iris Xe, Arc A-Series, Lunar Lake, Arrow Lake, Battlemage, and Ponte Vecchio
- ✓ Powers local LLM inference, swarm optimizers, and spiking neural network kernels
l0llm
Pure C# & Intel Level Zero native LLM runner and OpenAI-compatible REST server. 41.3 tok/s on Iris Xe with zero GC allocations.
Key Capabilities
- ✓ Zero Python, zero Conda, zero C++ framework dependencies — 100% managed C#
- ✓ Native support for LLaMA 3, Qwen 2.5, DeepSeek, Mistral, and Phi architectures
- ✓ Drop-in OpenAI REST API (/v1/chat/completions) with SSE streaming for Open-WebUI & LangChain
- ✓ Zero-allocation decode loop achieving 41.3 tokens/sec on mobile 15W Intel Iris Xe
l0check
Fast, standalone command-line diagnostic tool for inspecting Intel GPU architectures and validating compute kernels.
Key Capabilities
- ✓ Query driver versions, API versions, device IDs, PCI bus, and execution unit configurations
- ✓ Analyze memory subsystems, sub-device slices, and cache topology hierarchy
- ✓ Comprehensive diagnostic smoke tests and compute kernel pipeline verification
- ✓ Self-contained single-file executable with zero installation dependencies
ML.Agentic
Zero-dependency, in-memory building blocks for enterprise multi-agent systems and protocols on .NET 10.
Key Capabilities
- ✓ 21 canonical agentic patterns (Blackboard, HumanInTheLoopGate, ResourceBudgetLedger, Guardrails)
- ✓ Native protocol engine for Model Context Protocol (MCP), Agent2Agent (A2A), and JSON-RPC 2.0
- ✓ Sub-millisecond execution with 0 external package references or heavy graph stores
- ✓ Deterministic seeded testing and optional Level Zero GPU kernel offload
Poplar.NET
Enterprise .NET 10 runtime engine & resident graph orchestration SDK for Graphcore Colossus IPUs.
Key Capabilities
- ✓ Direct C# bindings and native C ABI shim for Graphcore Poplar & PopART SDK
- ✓ Pre-compiled on-chip resident graphs for LLM attention, swarm velocity, and covariance math
- ✓ Dynamic hardware telemetry, IPU fan governor, and zero-allocation FIFO streaming
- ✓ Paired with Poplar Explorer fork and IPU SQL Query Acceleration
Interactive PNN Decision Matrix
Interactive browser diagram of a probabilistic neural network decision matrix in HTML5/CSS/JavaScript.
Key Capabilities
- ✓ Live interactive neural decision matrix visualization in pure JavaScript/HTML5
- ✓ Dynamic parameter exploration and probabilistic decision boundary simulation
- ✓ Zero-install browser app hosted on GitHub Pages
- ✓ Full source code available on GitHub
Hypertensor GraphRAG
In-memory typed property-graph store with hybrid dense and BM25 retrieval, SIMD TensorPrimitives, and FastMCP agent context.
Key Capabilities
- ✓ In-memory typed property-graph store with dense & BM25 retrieval
- ✓ SIMD TensorPrimitives with optional Level Zero GPU offload
- ✓ 65+ tool FastMCP server & desktop HUD for agent context
- ✓ Tiered Open-Core and Commercial licensing
SwarmDev
A deterministic pipeline that turns a natural-language request into an inspectable plan with a diagnostic trail at every step.
Key Capabilities
- ✓ 6 explicit stages: Classify, Analyze, Graph, Route, Normalize, Emit
- ✓ Deterministic core with full regression testability
- ✓ Optional kGNN graph reasoning & HyperSwarm wavescan bridge
- ✓ Structured telemetry sidecar and hard policy interception gates
HFLabs Machine Learning Suite
Optimization, inference and agent reasoning in one .NET library set, with no cloud service in the loop.
Key Capabilities
- ✓ 5 HPO engines (Bayesian, Hyperband, BOHB, ASHA-BOHB, TuRBO) with JSON checkpointing
- ✓ 6 KGE models (TransE, RotatE, ComplEx, DistMult, ZPinch, Dual-Conj) & ontology extractors
- ✓ Level Zero LLM runtime with graph-constrained beam decoding and flash attention
- ✓ 21 ML.Agentic dependency-free primitives & cognitive orchestration loops
HyperSwarm Optimization Suite
Swarm optimization, from a single algorithm to a coordinated fleet. 35 optimizer modes with Intel Level Zero GPU kernels.
Key Capabilities
- ✓ 35 optimizers: 12 core swarm models, 14 hybrids, 9 advanced & multi-objective
- ✓ Director ensembles with dynamic budget & authority assignment
- ✓ 10 GPU offload classes with compiled Level Zero SPIR-V / Zebin kernels
- ✓ 123+ registered features & ~2,500 tests across 191 test suites
SwarmBench
Measure how swarms really behave, and watch them work. 34 optimization algorithms tested against 110 benchmark functions with real-time HUD visualizer.
Key Capabilities
- ✓ 110 registered test functions across 9 problem families (unimodal to high-D)
- ✓ SwarmBenchmarkRunner CLI covering streams, diagnostics & elimination tournaments
- ✓ SwarmVisualizer.Host desktop HUD with 6-layer frame telemetry & 7 lifecycle stages
- ✓ HyperCache config reuse & self-validating SwarmCharacterization LaTeX/JSON profiles
AutoMLApp (in development)
Design ML pipelines on a canvas. Run them on your hardware with 185+ node types across 25+ domains.
Key Capabilities
- ✓ 185+ node types across 25+ domains (Data, ONNX, IPU, SNN, Swarm, FPGA)
- ✓ Topologically sorted plan compilation with domain-specific step handlers
- ✓ F7 instant invariant validation (cycles, orphans, duplicate IDs, types)
- ✓ Pluggable HyperSwarm-backed AutoML & telemetry-aware cost scoring
Abigen (in development)
Point it at a C/C++ header, get reviewed, verified C# bindings — and keep them in sync as the native library moves.
Key Capabilities
- ✓ Narrow, fail-closed generator: binds top-level C functions, scalar fields & opaque handles
- ✓ Full manifest tracking with resolved libclang SHA-256 fingerprint & Clang args
- ✓ Release tracking for GitHub repos & build feeds with API word count diffs
- ✓ Desktop dashboard with 5-stage verification (smoke test, deterministic rerun, schema)
QueryAccel (MVP)
Offload PostgreSQL & SQL Server analytical queries to Graphcore IPUs with gRPC bridge microservice.
Key Capabilities
- ✓ PostgreSQL extension & SQL Server CLR bridge offloading aggregates and vector ops
- ✓ Hardware-accelerated COUNT, SUM, AVG, MIN, MAX, Cosine, Dot, L2 distance
- ✓ Integrated gRPC microservice connecting database engine to Poplar runtime
- ✓ Evaluated on Graphcore Bow-2000 and Bow Pod clusters