__ ______
/ / / / __ \
/ /_/ / /_/ /
/ __ / _, _/
/_/ /_/_/ |_|
Building Ainfera — an AI-Native model factory.
On the side: personal deployment packs for one box. Only Laguna is measured; receipts over hype.
I'm a founder who cares more about what actually runs than what trends.
- I build Ainfera — an AI-Native model factory (company work lives under @ainfera-ai).
- I publish personal, unaffiliated measurement packs so one DGX Spark stays reproducible.
- I optimize for evidence, clarity, and long-horizon craft — not launch theater.
Packs here are not company IP, not finetunes, and not invented numbers.
| Domain | What I actually do |
|---|---|
| Model systems | Serve paths, quant honesty, agentic runbooks, eval receipts |
| Inference craft | llama.cpp / GGUF pins, FP8 deployment previews, single-node MoE fit research |
| Agent loops | Tool-use harnesses, format/routing smoke, hermes-style checks |
| Evidence discipline | Digests, freeze clocks, SAQS claim binder, smoke ≠ headline |
| Product eng | TypeScript surfaces, Python toolchains, shell pack automation |
| Founder ops | Spec → measure → ship, ADHD-aware execution, conviction over consensus |
Personal deployment stack on one DGX Spark (GB10 · 128 GB). Only Laguna currently has family measurement authority.
Scripts · pins · cards · eval receipts.
Not finetunes. Not company IP. Not stretch claims.
| # | Pack | Role (honest) | Status |
|---|---|---|---|
| 01 | Laguna-S-2.1 · HF | Repo-maintenance deployment target · official Q4_K_M | Historical measured path · source candidate · freeze pending |
| 02 | Qwen3-Coder-Next · HF | Interactive coding · official FP8 | Preview · exact runtime model-load pending |
| 03 | DeepSeek-V4-Flash REAP25 · HF | Research pointer · REAP25/Pulsar reference | Docs-only HOLD / NO_HERO |
Size · pins · windows
Window (WIB)
List target 2026-08-03 12:00 · freeze target 2026-08-02 18:00
Release stays blocked until the August 2 freeze attestation passes across source, Hub, runtime safety, and claims.
Hero CTAs
Laguna Aug 3 20:00 only after explicit freeze clearance · Qwen null until a dated :8001 remeasurement and explicit gate · DeepSeek null (NO_HERO)
Laguna day-0 measured artifact
poolside/Laguna-S-2.1-GGUF · laguna-s-2.1-Q4_K_M.gguf
sha256 a8b55c75714ea73fd90ec85de5defdc0b8d88ca0ad2108343cdd8fc22f7583e4
engine pin 04b2b72 (poolsideai/llama.cpp, branch laguna) · measure tip bf82eab
format/routing smoke 40/40 · Hermes 27/27 validated-not-executed · ~21.47 t/s gen128
→ not long-horizon agent reliability proof · DFlash/NVFP4 not day-0 flagship
Official DeepSeek DSpark footprint reference
deepseek-ai/DeepSeek-V4-Flash-DSpark at revision 62af8fffb2f7030cac4de2f0169f5b8d1101b646: weight files total 166,886,535,336 bytes (155.425198 GiB). Repository-file accounting only; not a runtime measurement.
Upstream / quality refs (not day-0 serve authority)
poolside/Laguna-S-2.1-NVFP4 · Qwen/Qwen3-Coder-Next-FP8 · twaggs88 REAP25 DSpark GGUF · bases Laguna-S-2.1 · Qwen3-Coder-Next · DeepSeek-V4-Flash
measure → pin → publish receipts → refuse the stretch claim
| Principle | In practice |
|---|---|
| Conviction over consensus | One box. Real digests. Reproducible scripts. |
| Verify the premise | Smoke ≠ headline. Verifier ≠ gate clearance. |
| Honest labels | Experimental stays experimental until dated measure lands. |
| Integrity is structural | Claim binder ships with the pack — not a blog afterthought. |
Public claim binder for every Spark pack:
SPARK_AGENTIC_QUANT_STANDARD.md (SAQS)
Honesty locks — always on
diy_gguf = false— Laguna GGUF mirror hosts official Poolside bytes onlypublic_promo_before_launch = false- smoke ≠ agent headline · evidence-bind ≠ gate
- Qwen day-0 = official-FP8 Preview, not a throughput, quality, retention, or tool-success claim
- DeepSeek day-0 = docs-only, unmeasured HOLD / NO_HERO scaffold
- DSpark ≠ DGX Spark · peer GGUF ≠ the official DSpark artifact; listing is not a fit claim
- Packs are not Ainfera product surfaces, company eval, sales SKUs, or open finetunes
1 read README + INSTALL.yaml
2 pull the named upstream (or Laguna official GGUF + SHA256SUMS)
3 run Laguna only after its freeze + target receipt gates; run Qwen only as an explicit Preview; keep DeepSeek disabled
4 diff results/ — smoke ≠ headline
Static links only — no third-party stats widgets (those CDN pins break and show empty images).
| Track | Home |
|---|---|
| Personal packs | github.com/hizrianraz |
| Company | ainfera.ai · @ainfera-ai |
| linkedin.com/in/hizrian-raz | |
| Models | huggingface.co/hizrianraz |
| Profile binder | SPARK_AGENTIC_QUANT_STANDARD.md |
| Company | ainfera.ai · @ainfera-ai |
| linkedin.com/in/hizrian-raz | |
| Models | huggingface.co/hizrianraz |
| Pack questions | open a discussion on the relevant HF repo |
Thanks for reading this far.
Personal · unaffiliated measurement surface · not Ainfera product IP Polished 2026-07-30 · Laguna measure authoritybf82eab