Lossless and deterministic
Decompression reproduces the model bit-for-bit, every run and across releases — so the format itself contributes no error on top of the quantization. The accuracy you benchmark is the accuracy you ship.
Our Flagship Project
The Helios Project
Helios is our model-compression format. Its deployed configuration, HELIOS_H_M, reaches Q4_K_M-class quality at 57% of the file size — and every result here is measured on named hardware, not projected.
What it is
Big-model quality at a fraction of the footprint
Helios compresses a trained model into a compact container. The deployed configuration, HELIOS_H_M, matches the quality of llama.cpp's widely-used Q4_K_M quant at a little over half its size — so a model that needed 4.68 GiB fits in 2.69.
- The whole model compressed — nothing left at full precision
- One file, several memory/latency operating points
- Lossless: decompresses bit-for-bit, every run
The defining config
Q4_K_M-class quality at 57% of the file size
HELIOS_H_M, deployed · base model Qwen3-8B Instruct · WikiText-2 perplexity and a zero-shot suite of ARC-Challenge, WinoGrande, and PIQA
| Metric | HELIOS_H_M1 | Q4_K_M | Q2_K |
|---|---|---|---|
| File size | 2.69 GiB3.32 b/w | 4.68 GiB4.91 b/w | 3.06 GiB3.21 b/w |
| WikiText-2 PPLfp16 = 9.7152 | 10.0446+3.4% | 10.0831+3.8% | 10.8389+11.6% |
| Zero-shot trioacc / norm | 66.20 / 66.82 | 66.23 / 67.04 | 63.79 / 64.59 |
1 Deployed = the fully-compressed model, every component included — so the file size and bits-per-weight shown are the complete footprint a user actually gets, not the weights-only figure papers often quote (which ignores the rest of the model and looks smaller).
The balanced tier · HELIOS_M
Against Q2_K: better on every accuracy instrument, and smaller
| Metric | HELIOS_M | Q2_K | Result |
|---|---|---|---|
| File size | 2.40 GiB | 3.06 GiB | smaller |
| WikiText-2 PPL | 10.45 | 10.84 | 3.7% better |
| Zero-shot trio | 64.4–65.6 | 63.79 | better |
| Peak VRAM | 4.11 GiB | 4.09 GiB | parity |
VRAM, measured neutrally
Peak VRAM was read by a single external observer with both engines running as separate processes, each attributed on its own — neither engine reporting on itself. Helios lands at 4.11 GiB peak against Q2_K's 4.09. The full HELIOS_M stack runs at 3.84–3.85 GiB steady, with peak equal to steady — no transient overhead at all — and cold-starts in 7–8 seconds.
Holds on a second model
The result is not tuned to the base model. On Qwen2.5-7B the same approach beats Q2_K by 5.2% perplexity at a lower bit-rate (7.7535 vs 8.1808) and on the zero-shot trio (66.74 vs 66.08), and beats IQ2_M on every measure.
What it guarantees
Claims we can put numbers behind
One file, multiple operating points
A single file exposes several memory-versus-latency operating points, from a faster more memory-resident mode to a maximally compact one. You choose at load time, not at build time, and every mode returns identical output.
No training, no fine-tuning
Helios does not retrain or fine-tune the model to hit these numbers. Every benchmarked configuration comes from a 3–5 minute calibration — not a training run — so a model is ready in minutes, on the hardware you already have.
Speed, stated plainly
22.5–23.5 tok/s single-stream decode on an RTX 3090 (8B, 2048-token prompt); llama.cpp decodes the same model at 97–100. Our current runtime is at reading speed today, and a dedicated runtime built for throughput is in development — that is the honest state of this axis, not a number we hide.
Where it's headed
What's next
A dedicated runtime is in development to lift decode throughput well past today's reading-speed figure — the one axis where the current runtime trails llama.cpp. What you see here is the earliest deployed stage of what Helios can do.
“Our goal with Helios is to be the first step toward AGI — to accelerate the race to Artificial General Intelligence by making powerful models radically more efficient.”
For business inquiries regarding licensing of Helios
We are actively looking to license the Helios Project. Reach out to start the conversation.
Click here to get in touch