---

Who we are

A one-person, non-profit model lab. We start from strong open-source bases — first up, DavidAU's Qwen3.8-27B Cold Fusion — push them as fast and as smart as possible with current speed tactics, then release every quant, free, under Apache-2.0.

That's the whole thing. Nothing is for sale.

Mission

Same brains. More tokens per second. Proven, not vibed.

• Fast — the toolkit we're aiming at: speculative decoding (draft + EAGLE-3), NEO-imatrix int4, flash-attention, reasoning-effort compression. Goal is a real multiple on tok/s with outputs unchanged — measured, not assumed.
• Smart — every release will ship a tok/s table and a quality A/B vs the base. No "smarter than base" without a measured A/B.
• Open — all quants, all use cases, one gloriously cursed card at a time.

Flagship

First target: DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 (Apache-2.0) — a 27B dense, vision-language model with 64 layers, 16 full-attention layers (4 KV heads), MTP self-speculation, and 262K native context (→1M extensible). The plan is to re-quantise it (NEO-imatrix) and speed it up (draft/EAGLE-3 speculative), then publish an honest A/B.

No benchmarks here yet — nothing has been trained or measured. Speed and quality numbers will land on each release card once they come from our own runs. The full recipe and blank result tables live in FINETUNE-PLAN.md.

The Fleet

Rig is being assembled — the 170HX host isn't in hand yet.

Stack: Ubuntu 22.04 · CUDA 12.x (never 13 — Turing + GA100 are dropped there) · 570+ open-kernel driver (for the 170HX unlock) · llama.cpp (MTP + imatrix) · vLLM on Ampere. The 170HX lives on its own pinned host — snapshot + apt-mark hold, never shares a driver stack.

How we cook

1. Baseline the base — fixed prompts, tok/s + a small quality set. Every later number is measured against this.
2. Free-lunch config — flash-attn, -ctk/-ctv q8_0, NEO-imatrix int4, tuned low/medium reasoning_effort.
3. Draft-model speculative in llama.cpp — pair a Qwen3-0.6B draft with the 27B. Lossless, headline speed-up.
4. Stretch: EAGLE-3 head on vLLM/SGLang — bigger acceptance, same output distribution. The signature move nobody else trains.

Operating principles

1. Proof before panic. Measure, don't vibe.
2. Prove it's lossless — or say what the speed cost.
3. Boring-reliable beats clever-fragile. The 170HX stays pinned.
4. Ship the series, not the single. One legendary model beats twenty random merges.

Goals

• Acquire + unlock the 170HX host, pin the CUDA 12.x / driver stack
• Baseline the target model on our own hardware (tok/s + quality set)
• NEO-imatrix quant chain
• Draft-model speculative decoding (llama.cpp)
• EAGLE-3 speculative head — same outputs, faster (target TBD by measurement)
• Benchmark harness with honest A/B vs base on every release
• First public release, receipts on the card

---