100.0% Native Safe Rust + Native Apple Silicon Metal GPU · 0.0% Python · 0 External Crates · Tri-OS LLM Precision Studio
Drag-and-Drop Mixed-Precision Quantizer (Bonsai 1-Bit + Gyro-Coder MoE Pruning) · Multi-Agent Asymmetric KV Planner · Streaming Abliteration · Static Multi-LoRA Merger · NALA-RAG & 28-Tool MCP-cORe
🇩🇪 Deutsch · 🇬🇧 English · 🇫🇷 Français · 🇮🇹 Italiano · 🇨🇭 Rumantsch · 🪨 Caveman Install · ❓ FAQ (5 Lang)
PDF Guides: DE.pdf · EN.pdf · FR.pdf · IT.pdf · RM.pdf
| Engine / Module | Measured Hardware Result | Integrity & Security |
|---|---|---|
RSI-MF-05 Apple Metal GPU (nala-modelforge-metal) |
26.5× – 28.7× Speedup, 6.66 GWeights/s, 27.66 GB/s (16.7M weights in 2.52 ms on Apple M4 Max Zero-Copy Shared Memory) |
0 bit mismatches, max_scale_drift = 1e-7, 0% Python |
RSI-MF-01 Parallel-SIMD + MSL/GLSL Shaders |
4.45× Multi-Core Speedup, 476.6 MWeights/s | 0.000000 bit drift, embedded SHA-256 verified shaders |
RSI-MF-02 Gyro-Coder Repo & MoE Expert Profiler |
Scans repo in 7.69 ms, prunes 512 → Top-256 Coding Experts (-46% VRAM) |
Read-Only, skips .git/.env/*.pem/*.key/secrets |
RSI-MF-03 & RSI-MF-06 Streaming GGUF Abliteration |
4 MB chunk buffer (< 16 MB RAM even on 1.5 TB GGUF), F32/F16/Q8_0/Q4_0, >98% Refusal Suppression, >0.995 Coding Cosine |
Atomic .tmp → rename Guard (source SHA-256 100% untouched) |
RSI-MF-07 Static Multi-LoRA GGUF Merger |
0.0 ms inference adapter overhead (W + Σ(α/r)·B·A) |
Frobenius Spectral-Drift Guard + atomic rollback |
RSI-MF-08 Direct NALA-RAG (:18700) Dataset Builder |
Exports imatrix.txt + coding_lora.jsonl + SHA256SUMS.txt in 0.61 ms |
5-Zone Tenant Firewall + RFC-1918/IBAN/Email Zero-PII Redactor |
Pure-Rust Deslop Gate (--deslop-audit) |
5/5 PASS in 54–101 µs |
10× faster than Python, 0.0% Python anywhere in repo |
- Zero-Copy Streaming Binary GGUF (v2/v3) & Safetensors Inspector: Drag any local
.gguf,.safetensors, or HuggingFace directory directly into the macOS AppKit window (NSDraggingDestination) — inspects headers in< 25 mswith zero RAM copy even for 1.5 TB MoE models. - Domain-Selective Mixed-Precision Auto-Fit: Preserves Coding, Logic, Architecture, Tool-Calling, Attention (
attn_q/k/v/o), Embeddings & FP16 Router Gates at maximum precision (BF16/Q8_0/Q6_K) while compressing secondary weights via Randomized Walsh-Hadamard Transform (H·D·x) to Bonsai 1-Bitg128(1.125 bpw) or BitNet 1.58-Bit Ternary. - Asymmetric Multi-User & Multi-Agent KV-Cache Planner: Pre-configured for heterogeneous swarms (e.g.,
1× Human 1M+1× Orchestrator 1M+14× Specialist Agents 256K) with GQA/MLA/Hybrid-GDN math, Radix Prefix Sharing, YaRN RoPE scaling, and NALA-Strata-RS NVMe Disk-Parking. - Multi-LoRA Stacking +
NALA-RAG (:18700)+NALA-MCP-cORe(28 Tools): Native LoRA file picker, spectral-drift verifier, static GGUF merger, live NALA-RAG Evidence-Lock bridge, opt-in Zero-PII NALA-Grid 4D HMAC-SHA256 orchestrator, and built-in--mcp-stdioJSON-RPC 2.0 server. - Tri-OS Native Distribution: macOS ARM64 & Intel
.dmg/.app, Ubuntu/Debianamd64&arm64static musl.deb/.tar.gz, Windows 11x86_64.exe/.zip(+ GitHub Actions Windows ARM64 CI), CycloneDX 1.5 SBOM + SLSA v1.0, andSHA256SUMS.txt.
# Launch Studio UI on 127.0.0.1:18795
nala-modelforge-rs --open
# Wire all 28 MCP tools into NALA-MCP-cORe, Antigravity, Gemini CLI & ZCode (Zero-Python)
nala-modelforge-rs --register-mcp
# Run live Apple Silicon Metal GPU 1-Bit Bonsai benchmark (RSI-MF-05)
nala-modelforge-rs --rsi-mf05
# Verify all 12 math, security, and Golden-DOM gates
nala-modelforge-rs --self-test