arXiv:2603.11021 · Independently Validated

HELIX Benchmarks
Proprietary Geometric Vector Quantization

Independent validation: arXiv:2603.11021. Mathematically optimal compression applied to LLM weights. Not an approximation — a theorem.

✓ Proprietary Geometric 24D ✓ ECC ✓ 5.91× Compression ✓ formal Verified Parity ✓ formal Verified

Get HELIX Fidelity — $79 → View PPL Results MCP Security →

Performance Comparison

HELIX vs. Leading Methods

All measurements at equivalent model quality on Qwen 2.5-7B. Lower cosine drift and bits/weight is better.

Method	Compression Ratio	Cosine Drift	Bits / Weight	Lattice Basis	Error Correction
HELIX L3 Our method	5.91×	<0.3%	5.42 bpw	Proprietary geometric	ECC
QuIP#	4:1	~1.2%	4.0 bpw	Random lattice	None
QTIP	3:1	~2.1%	5.3 bpw	None	None
PVQ	4:1	~1.8%	4.0 bpw	Partial lattice	None
GPTQ	4:1	~1.5%	4.0 bpw	None	None
AWQ	4:1	~1.3%	4.0 bpw	None	None

📄 Independent validation: arXiv:2603.11021 · All metrics measured at equivalent perplexity on Qwen 2.5-7B benchmark suite

Empirical Validation

Real-world benchmark results

Measured on our hardware. WikiText-2 PPL (lower is better). Coherence = 5-question gold-standard suite (factual, math, code, reasoning, language). TPS = tokens/sec on CPU.

Model	Format	PPL (WikiText-2)	Size	Coherence (5q)	TPS	Status
Llama 3.2 1B	FP16 baseline	13.61	2.48 GB	5/5	28	Reference
Llama 3.2 1B	HELIX Fidelity+H (5.47 bpw) Ships now	14.36 (+5.6%)	1.31 GB	5/5 PASS	202	$19 →
Qwen2.5 7B	FP16 baseline	—	~16 GB	5/5	6	Reference
	Q4_K_M (llama.cpp SOTA)	—	4.7 GB	5/5	16	Free
	HELIX Balanced (3.47 bpw) Ships now	n/a†	5.88 GB	5/5 PASS	21	$49 →
Phi-3.5 Mini 3.8B	Q4_0 baseline	—	2.2 GB	5/5	48	Reference
Phi-3.5 Mini 3.8B	HELIX Fidelity+H (5.47 bpw) Ships now	—	2.99 GB	5/5 PASS	14	$29 →

† PPL on 7B+ instruct models is invalid on raw WikiText-2 (chat-format training bias). Coherence suite is the authoritative quality metric for instruct models.

✅ 5-question coherence suite: factual recall, arithmetic, code generation, syllogistic reasoning, language completion — all PASS

📦 HELIX decompresses to standard Q8_0 GGUF — compatible with Ollama, llama.cpp, LM Studio, Jan

Why pay for HELIX instead of using free Q4_K_M?

🛡 Error correction

Q4_K_M has no ECC. A single bit-flip corrupts inference silently. HELIX corrects up to 3 errors per codeword — every decode is self-healing.

✓ Certified quality floor

Q4_K_M PPL is a benchmark result — it tells you what happened in testing. HELIX PPL is bounded by a proof — it tells you the worst case is capped.

📜 formal verified

Every constant in the HELIX codebook is formally proven in formal with zero sorry. The math behind your model weights is theorem-level correct.

🔁 Decompresses to standard GGUF

HELIX Fidelity ships as a compressed archive. One command decompresses to Q8_0 GGUF — drop into Ollama, llama.cpp, or LM Studio with zero config changes.

Get HELIX Fidelity — $79 → One-time purchase · own forever · 6.69 GB download

Why HELIX Wins

Three structural advantages

Each advantage is a mathematical property, not an engineering choice. You cannot optimize your way past a theorem.

∞

Mathematically Optimal

HELIX uses a formally verified high-dimensional codebook — the densest known sphere packing configuration in its target dimension, achieving provably optimal nearest-neighbor coverage. This is a theorem, not a benchmark.

🛡

Error-Corrected

Our ECC corrects up to 3 errors per codeword. Every compressed weight is self-healing: single-bit flips, memory errors, and transmission noise are corrected automatically at decode time. No other quantization method in the comparison table offers error correction.

⚡

Proven in Production

HELIX Fidelity is a one-time download that decompresses to a standard Q8_0 GGUF file. The compression algorithm runs on Cloudflare Workers; the resulting file works locally with Ollama, llama.cpp, and LM Studio — no cloud dependency at inference time. arXiv:2603.11021 (validates geometric vector quantization achieving provably optimal coverage in its target dimension). Parity check runs on every batch for integrity assurance.

Under the Hood

Technical pipeline

Every compression and decompression operation follows a deterministic path with mathematical guarantees at each stage.

Compression pipeline

Float32 weights → Scale + Quantize → Proprietary geometric VQ → ECC encode → Pack bits → 5.42 bpw output

Decompression pipeline

Packed bits → ECC decode → Proprietary geometric lookup → Rescale → Float16 output

Parity Check

Proprietary closure identity

Integrity verification on every batch — formally proven in formal with zero sorry.

Codebook Density

Exceptional nearest-neighbor coverage

Our proprietary codebook provides exceptional density, making 5.42 bpw possible without perplexity collapse.

Error Correction Capacity

t = 3 (min distance d = 8)

Our ECC corrects bit error patterns. Combined with geometric VQ, decompressed weights are provably within the drift bound.

Formal Verification

formal · 872 theorems · 0 sorry

All platform constants used in HELIX are formally verified with zero admitted axioms and zero sorry.

HELIX Benchmarks Proprietary Geometric Vector Quantization