Roadmap
v1.0.0 — Stable Release (2026-07-25)
Completed
All 8 development phases complete, API frozen for stability, full regression suite passes:
| Area |
Deliverable |
Status |
| Model Zoo |
from_pretrained(), ModelHub, ModelConfig/Card/Registry, C/C++/Python integration |
Done |
| Distributed Training |
ZeRO-1/2/3, pipeline/tensor/expert parallel, elastic, fault tolerance |
Done |
| Advanced Architectures |
Differential Attention, Mamba-2, FlexAttention, MoD, gated activations, YaRN, ALiBi |
Done |
| Security |
S0-S9 complete (21,984+ LOC), 4-ring firewall, 123 test executables |
Done |
| Quantization |
INT8/FP8/AWQ/GPTQ, Python + CUDA kernels |
Done |
| Checkpointing |
Async save/load, heartbeat, elastic training, binary format reader |
Done |
| Profiling |
Profiler, structured logging, NVTX, sanitizer scripts |
Done |
| CUDA Backend |
Tensor-core GEMM, Flash Attention v2/v3, autodiff, memory pool, RNG |
Done |
v1.1 — Model Serving, Fine-Tuning & Model Hub (Completed — v1.1.1)
| Area |
Deliverable |
Status |
| Serving |
C HTTP Server, FastAPI inference engine, continuous batching |
Done (v1.1.1) |
| Fine-tuning |
LoRA/QLoRA adapters, DPO/GRPO trainers |
Done (v1.1.0) |
| Evaluation |
Synthetic and benchmark eval harnesses (MMLU, GSM8K, HumanEval) |
Done (v1.1.0) |
| Model Hub |
Centralized registry, CLI (sneppx-hub), storage backend, C WinHTTP client |
Done (v1.1.1) |
| Security Hardening |
C safety refactoring, Dilithium/Kyber crypto fixes, Ninja build presets |
Done (v1.1.1) |
| ONNX toolkit |
Standalone numpy-only parse/serialize/check/infer/optimize/QDQ/external-data, numpy + onnxruntime runtimes, sneppx-onnx CLI, legacy-wire-compatible exporter bridge |
Done (Unreleased) |
v1.2 — Production Hardening
| Area |
Deliverable |
Metric |
| Performance |
CUDA graph capture, INT4/FP8 inference kernels |
2x inference speedup |
| Memory |
PagedAttention v2, vLLM-style KV cache |
4x batch size increase |
| Reliability |
Circuit breakers, retry logic, health checks for distributed training |
99.9% job success |
| Monitoring |
Prometheus metrics, Grafana dashboards, tracing |
Real-time visibility |
v2.0 — Competitive with LLaMA-3
- 70B parameters, 1M LOC
- Multi-node training, 64+ GPUs
- Adaptive state size per HSS layer
- Hierarchical MoE with 256 experts
- Self-modifying NPE programs with meta-optimization
- Cross-datacenter federated learning
- S4 ZK proofs complete, S5 on-device runtime
- Formal verification of critical paths
v3.0 — Competitive with GPT-4
- 1T parameters, 5M LOC
- Fully learned routing, 1000+ GPU cluster
- S6 federated contribution protocol
- On-device attestation
v4.0 — Beyond Existing Systems
- 10T parameters, 15M LOC
- Self-evolving architecture
- S7 self-evolving NPE paths
- S8-S9 formal verification + pentest
- Full formal proof of alignment
Quarterly Milestones (Updated)
| Quarter |
Milestone |
Status |
| 2026 Q3-Q4 |
Phase 1-8 complete, v1.0.0 release |
Done |
| 2027 Q1 |
Model serving (vLLM/TensorRT-LLM), LoRA fine-tuning, LM Eval Harness, FSDP |
✅ Done |
| 2027 Q2 |
CUDA graph capture, PagedAttention v2, Prometheus monitoring |
Pending |
| 2027 Q3 |
7B model pre-training, distributed checkpoint resume |
Pending |
| 2027 Q4 |
Multi-GPU training, FSDP, 13B model |
Pending |
| 2028 |
70B model, S6-S9 full integration, formal verification |
Pending |