PALMFUTURE_

PALMFUTURE.SPACE // MISSION FILE

WISIT SRIMALA

Tinkerer · Self-hoster · Local-AI enthusiast

LOC // Bangkok, ThailandUTC+7
INITIALIZEVIEW LOG
COORD 14.0°N 100.5°ECLK Sun 13:04:34SYS // ONLINESIGNAL ♥ 100%
// 01 · ABOUT
palmfuture@homelab: ~
calibrating skill modules · 8 online
0%
Operator file_
Tinkerer, self-hoster, and local-AI enthusiast. I run a small AI lab on the desk beside me, quantize open models in my spare time, and print the parts I need on a Bambu Lab A1 I can't stop upgrading. Based in Bangkok, Thailand.
  • // AI — 2× MSI EdgeXpert (DGX Spark), DeepSeek V4 Flash · TP2
  • // QUANT — GPTQ / FP8 / NVFP4 model packaging
  • // INFRA — Docker, Nginx, Go self-hosted grid
  • // FAB — Bambu Lab A1, PETG
// 02 · SYSTEMS
MOD 01GO-LANG
Go
LVL 80/100SERVICES / CLIS
MOD 02TS-REACT
TypeScript / React
LVL 75/100NEXT.JS
MOD 03NODE
Node.js
LVL 70/100JS ECOSYSTEM
MOD 04PYTHON
Python
LVL 72/100SCRIPTING / ML
MOD 05DOCKER-HOMELAB
Docker / Homelab
LVL 85/100SELF-HOSTED
MOD 06LOCAL-LLM
Local LLM Ops
LVL 78/1002× DGX SPARK · TP2
MOD 07QUANT
Quantization
LVL 82/100GPTQ · FP8 · NVFP4
MOD 083D-PRINT
3D Printing
LVL 60/100BAMBU LAB A1 · PETG
// 03 · PROJECTS
PRJ / 01STABLE
Qwen3.6-35B-A3B-GPTQ-Int4
Flagship quant: a 35B MoE repackaged to GPTQ-Int4. ~580k downloads, 29 likes, and small enough to run on consumer hardware.
GPTQ-INT435B MOEQUANT
huggingface.co / Qwen3.6-35B-A3B-GPTQ-Int4
PRJ / 02ACTIVE
Qwen3.8-27B-GPTQ-Int4
NEW
Newest quant (Aug 2026): Qwen3.8-27B dense → GPTQ-Int4, uniform g32, MTP speculative decoding + vision, vLLM-verified.
GPTQ-INT427B DENSEVISION+MTP
huggingface.co / Qwen3.8-27B-GPTQ-Int4
PRJ / 03STABLE
Qwen3.6-27B-GPTQ-Int4
A dense 27B repackaged to GPTQ-Int4. 24,272 downloads, 1 like — another quant that runs quietly on consumer hardware.
GPTQ-INT427B DENSEQUANT
huggingface.co / Qwen3.6-27B-GPTQ-Int4
PRJ / 04ACTIVE
vllm-default-thinking-budget
Injects a default thinking_token_budget and presence_penalty into vLLM so models keep their reasoning behaviour out of the box.
VLLMSHELLREASONING
github.com/palmfuture/vllm-default-thinking-budget
PRJ / 05ACTIVE
Ornith-1.0-35B-NVFP4A16
A 35B repackaged to NVFP4-A16. 1,536 downloads, 2 likes — part of the NVFP4 line for edge-sized serving.
NVFP435BQUANT
huggingface.co / Ornith-1.0-35B-NVFP4A16
PRJ / 06EXPERIMENTAL
Nex-N2-mini-NVFP4A16
A compact N2-mini repackaged to NVFP4-A16. 129 downloads — a small everyday model for light tasks.
NVFP4MINIQUANT
huggingface.co / Nex-N2-mini-NVFP4A16
PRJ / 07EXPERIMENTAL
gemma-4-12B-it-FP8
Gemma-4 12B instruct repackaged to FP8. 11 downloads, 1 like — the lightweight end of the quant line.
FP812BQUANT
huggingface.co / gemma-4-12B-it-FP8
// 04 · OPS LOG
  1. 2025 · Q2
    HOMELAB ONLINE
    An old machine gets promoted to an always-on server in Bangkok. First self-hosted service survives a month without intervention.
  2. 2025 · NOV
    RTX 3060 ×4 CLUSTER
    A four-card RTX 3060 cluster comes online for local GPU work — the first real compute, before anything more serious arrived.
  3. 2025 · DEC
    LOCAL LLM GOES LIVE
    The desk starts serving open models locally instead of renting GPU calls for small jobs.
  4. 2026 · FEB
    QUANTIZATION SHIPS
    Open weights get repackaged into GPTQ / FP8 / NVFP4 and published on Hugging Face — now crossing 300k total downloads.
  5. 2026 · APR
    DGX SPARK ONLINE
    An MSI EdgeXpert cluster takes over local model serving with DeepSeek V4 Flash in TP2.
  6. 2026 · JUN
    FIRST PRINTER
    A Bambu Lab A1 arrives. Calibration cubes first, then parts that needed to exist.
  7. 2026 · AUG
    QWEN3.8-27B QUANT SHIPS
    Qwen3.8-27B dense gets repackaged to GPTQ-Int4 (uniform g32) on the 4× RTX 3060 cluster, with MTP speculative decoding and vision verified on vLLM.
// 05 · LINK
ESTABLISH
LINK
UPLINK // EMAIL
[email protected]
github.com/palmfuturehuggingface.co/palmfuturelinkedin.com/in/wisit-srimala
Currently routing signals from BangkokUTC+7