# ethanyeang — full site context

> Independent developer in Shenzhen working on Physical AI, robotics, AI agents and creative engineering. Product definition and business lead who also works hands-on with software, model inference and hardware validation alongside a technical team.

Canonical site: https://ethanyeang.me/
Languages: zh-CN and en (reciprocal hreflang)
Content updated: 2026-11-04

## Who this site is for
- Technology enthusiasts and hardware tinkerers: Physical AI, robot experiments and public datasets.
- Gamers, handheld and DIY hardware fans: R36S hardware guides and local large-model rig notes.
- People who value craft and design: original photography and a curated project archive.
- Teams: custom AI agents, private deployment, robotics integration and prototypes.
- Angel investors and early-stage partners: project evidence, open data and direct contact.

## Pages
- Home: https://ethanyeang.me/ (en: https://ethanyeang.me/en/)
- Projects: https://ethanyeang.me/projects/ (en: https://ethanyeang.me/en/projects/)
- About and collaboration: https://ethanyeang.me/about/ (en: https://ethanyeang.me/en/about/)
- About: https://ethanyeang.me/about/ (en: https://ethanyeang.me/en/about/)
- Field notes: https://ethanyeang.me/field-notes/ (en: https://ethanyeang.me/en/field-notes/)
- Models and datasets: https://ethanyeang.me/lab/ (en: https://ethanyeang.me/en/lab/)
- Contact: https://ethanyeang.me/contact/ (en: https://ethanyeang.me/en/contact/)

## Services
Enterprise AI agents and private deployment; Physical AI, robot integration and VLA validation; product prototypes and custom software. Work starts from a concrete business goal, with testable scope agreed before prototyping. Chinese and English; Shenzhen on-site or remote. Contact: yzc_2024@163.com

### Service FAQ
- Q: What kinds of projects do you take on?
  A: Enterprise agents and private deployment, robot integration and real-world VLA validation, and product prototypes with custom software. We start from the goal and constraints, then agree on an assessment, prototype or full delivery.
- Q: Do you work remotely?
  A: Yes. Shenzhen projects can involve on-site meetings and robot testing; other locations work mostly remotely, in Chinese or English.
- Q: How does a project usually start?
  A: Send an email with context, goal, existing hardware or data, expected timeline and budget range. The first conversation defines the problem and boundary, then we agree stages and deliverables.
- Q: How do you scope and price the work?
  A: Pricing follows scope, hardware, data readiness and delivery format. Exploration, prototype and full delivery are usually separate stages, each confirmed before the next.
- Q: Which robot platforms are supported?
  A: SO-101, B601, Reachy Mini, LeKiwi and common LeRobot / VLA workflows can be evaluated. Actual support depends on hardware, firmware, drivers and site conditions; uncertain parts are validated at small scale first.
- Q: Do you guarantee model performance or success rates?
  A: No promises about unverified metrics. We define testable acceptance criteria first and rely on real data and field results. Site demos document specific experiments.

## Contact and profiles
- Email: yzc_2024@163.com
- GitHub: https://github.com/metahubaifeel
- Hugging Face: https://huggingface.co/userresu2026
- X: https://x.com/ethanyeang
- Xiaohongshu: https://www.xiaohongshu.com/user/profile/42277431340

## Project archive

### B601 × LingBot-VLA v2 / B601 × LingBot-VLA v2
- zh: https://ethanyeang.me/projects/b601-lingbot-vla2/
- en: https://ethanyeang.me/en/projects/b601-lingbot-vla2/
- category: robotics; type: 实机推理
- topics: B601, LingBot-VLA v2, LeRobot, Physical AI
- en summary: Bringing teleoperation demos, VLA inference and real dual arm manipulation together on the B601.
- en notes: This project connects teleoperated demonstration collection, a dual arm LeRobot interface and LingBot-VLA v2 inference on a B601 platform. Selected footage shows tabletop pick, move and stack experiments. The clips document specific experiments; they **do not imply universal success rates or general autonomous capabilities**.
- source: https://huggingface.co/datasets/userresu2026/b601-lingbot-vla2
- video: https://ethanyeang.me/assets/video/b601-demo.mp4
- video: https://ethanyeang.me/assets/video/b601-data-collection-full.mp4 (B601 · One unbroken collection run)
- video: https://ethanyeang.me/assets/video/b601-overview.mp4 (Project overview · cut and continuous runs)
- video: https://ethanyeang.me/assets/video/b601-data-collection.mp4 (B601 · Real-world data collection)
- video: https://ethanyeang.me/assets/video/b601-wrist-camera-mount.mp4 (B601 · Wrist camera mount prototype)
- video: https://ethanyeang.me/assets/video/lekiwi-bubble.mp4 (LeKiwi × LingBot-VLA v2 · bubbles)
- public data: b601-dm-20260718_0031 https://huggingface.co/datasets/userresu2026/b601-dm-20260718_0031; b601-dm-20260718_0053 https://huggingface.co/datasets/userresu2026/b601-dm-20260718_0053; b601-dm-icehockey-smoke https://huggingface.co/datasets/userresu2026/b601-dm-icehockey-smoke; b601-dm-icehockey-v1 https://huggingface.co/datasets/userresu2026/b601-dm-icehockey-v1; b601-lingbot-vla2 https://huggingface.co/datasets/userresu2026/b601-lingbot-vla2

### T800 / Backflip Retargeting / T800 / 后空翻动作重定向
- zh: https://ethanyeang.me/projects/t800-backflip-sim/
- en: https://ethanyeang.me/en/projects/t800-backflip-sim/
- category: robotics; type: 动作重定向 · 仿真
- topics: Retargeting, IK, MuJoCo, Kinematic replay, mjlab
- en summary: Moving G1 backflips from an open motion dataset onto the T800 humanoid: 1.227× leg-length retargeting, per-frame IK correction, and MuJoCo kinematic replay rendered at 50 fps. The RL branch failed to produce a policy.
- en notes: Retargets Unitree G1 backflips from the UniLab dataset onto the T800 humanoid: pelvis trajectory scaled by the 1.227 leg-length ratio, per-frame IK correction with scipy.optimize.least_squares, then kinematic replay and offscreen MuJoCo rendering into 50 fps landscape and portrait clips. The motion comes from the reference trajectory, not from a policy; the RL vault training did not converge. The page also carries a few locomotion clips: real-machine ramp walking at original speed, and a simulation standing up from lying down, front and back. They record what was on camera; they do not show that a policy works.
- video: https://ethanyeang.me/assets/video/t800/flip360.mp4
- video: https://ethanyeang.me/assets/video/t800-uphill-real.mp4 (Real machine · up the ramp)
- video: https://ethanyeang.me/assets/video/t800-sim-standup.mp4 (Simulation · standing up, front then back)
- video: https://ethanyeang.me/assets/video/t800-policy-vs-ref.mp4 (Policy vs reference · T800 walking)

### MicroDuck / 牛来 / MicroDuck / 牛来
- zh: https://ethanyeang.me/projects/microduck-jump-rl/
- en: https://ethanyeang.me/en/projects/microduck-jump-rl/
- category: robotics; type: 强化学习 · 仿真
- topics: PPO, MuJoCo, ONNX, Simulation
- en summary: From crouching to hopping: designing rewards that teach a small biped to jump in simulation.
- en notes: This project explores PPO reward design and motion stability in a simulation training environment for MicroDuck, an open source robot from Pollen Robotics and Hugging Face. Policies are exported to ONNX. The video shows **simulation results, not a claim of real robot validation**. Added on 2026-10-31: the Thor training run and a batch of MicroDucks stepping in parallel environments.
- source: https://github.com/metahubaifeel/microduck-jump-rl
- video: https://ethanyeang.me/assets/video/microduck.mp4
- video: https://ethanyeang.me/assets/video/microduck-rl-walk.mp4 (119 seconds of training on Thor)
- video: https://ethanyeang.me/assets/video/microduck-parallel-envs.mp4 (Parallel environments · 8 seconds)

### Q1 / Robot Console / Q1 / Robot Console
- zh: https://ethanyeang.me/projects/q1-console/
- en: https://ethanyeang.me/en/projects/q1-console/
- category: robotics; type: 机器人交互
- topics: Web UI, Robot Control, Camera, Voice
- en summary: The robot's stereo camera, telemetry, voice and motion control in one browser console — operating a Yuan Q1 humanoid in Shanghai from Shenzhen.
- en notes: A browser console for the Yuan Q1 / T1 humanoid that keeps its state and controls in one workspace: the stereo camera feed, battery and attitude instruments, speed and odometry, posture modes with WASD teleoperation, map and navigation, plus voice, expression lighting and hardware diagnostics. The camera is the robot's own pair of eyes, not a head-mounted add-on — the picture is the two views it actually has.

## Tested in Shanghai, driven from Shenzhen

After getting test access in September I ran the first round on the robot in Shanghai. Before leaving I set up the remote link; back in Shenzhen I operated that same robot through the same console. Nobody was in the room, the picture came from its stereo camera, and round-trip latency stayed **under 300 ms** — quick enough to actually drive. Remote operation covers far more situations than a one-off demo, which is why I want to keep working on it.

## About the recording

The recording below is a live session on 14 September 2026, captured entirely from the console itself. Capabilities vary with firmware and hardware; check the repository for compatibility and installation.
- source: https://github.com/metahubaifeel/q1-console
- video: https://ethanyeang.me/assets/video/q1/q1-console-overview.mp4
- video: https://ethanyeang.me/assets/video/q1/q1-console-camera.mp4 (Stereo camera · live view)
- video: https://ethanyeang.me/assets/video/q1/q1-console-control.mp4 (Motion control · safety unlock)

### BotverseX / BotverseX
- zh: https://ethanyeang.me/projects/botversex/
- en: https://ethanyeang.me/en/projects/botversex/
- category: robotics; type: 具身智能工作台
- topics: BotverseX, Physical AI, LeRobot, Teleoperation
- en summary: BotverseX is my embodied AI workbench, connecting robots, teleoperation, data recording, training and deployment.
- en notes: The long view is a hardware-focused companion to Hugging Face: one workbench where a maker can walk the same path — **connect, calibrate, teleoperate, record, train, deploy, infer**. Current work spans physical SO-101 and B601 robots as well as MicroDuck simulation. Support differs by platform; some workflows require a bridge or dedicated drivers.

The point cloud on the homepage is not a video: it is the **robot body sampled from its URDF**, streamed live and following the pointer. Change the machine and the body changes with it — no client install, feel first.
- source: https://botversex.dev
- video: https://ethanyeang.me/assets/video/botversex-site-loop.mp4

### World Model Playground / World Model Playground
- zh: https://ethanyeang.me/projects/world-model-playground/
- en: https://ethanyeang.me/en/projects/world-model-playground/
- category: experiments; type: 三维重建实验
- topics: 3D Reconstruction, Spatial, Experiment
- en summary: Reconstructing a makerspace scene and exploring camera views and browsable 3D environments.
- en notes: This local experiment preserves rendered views of a makerspace scene. It is a **visual exploration of reconstruction and spatial tools, not a claim about a general purpose world model**.
- video: https://ethanyeang.me/assets/video/world-model.mp4

### R36S / r36s.space / R36S / r36s.space
- zh: https://ethanyeang.me/projects/r36s-space/
- en: https://ethanyeang.me/en/projects/r36s-space/
- category: products; type: 独立站 · Journal
- topics: R36S, Independent Store, Technical Journal, Open Hardware
- en summary: An independent R36S handheld shop and technical journal covering board identification, setup, SD cards and save files.
- en notes: I run r36s.space, an independent bare-handheld shop and technical journal. The journal documents board revisions, setup, SD cards, save files and buying decisions, **all checked against hardware I actually took apart**.

Product pages describe **exactly what is in the box and bundle no games**. Read the original articles on r36s.space; related code projects remain linked to GitHub.
- source: https://r36s.space
- video: https://ethanyeang.me/assets/video/r36s-shoulder-buttons.mp4 (R36S · Shoulder button modification)
- video: https://ethanyeang.me/assets/video/r36s-3d-grip.mp4 (R36S · 3D-printed grip prototype)

### Handheld Tinkering / 掌机改造
- zh: https://ethanyeang.me/projects/handheld-tinkering/
- en: https://ethanyeang.me/en/projects/handheld-tinkering/
- category: experiments; type: Handheld · Linux · Retro Gaming
- topics: R36S, RG353V, ROCKNIX, ArkOS, Retro Gaming
- en summary: Two handhelds, four Linux builds and a 3D-printed grip: taking the R36S and RG353V apart until they feel like machines worth holding.
- en notes: I started with handhelds last September and could not tell a “clone handheld” from an Android TV box. Then I flashed one system after another — EEOS, ArkOS, RockNIX, DarkOS — and realised they are all trimmed-down Linux. The fun moved from playing games to changing the machine.

## Two machines, both in use

Two devices today. The R36S got a **3D-printed grip, a shoulder-button modification and a new back shell**, and finally feels built to be held; the RG353V was opened to look at the board, repartitioned with fdisk, given its own boot scripts, and now sits under the TV as a small console with a controller. The games still matter: the 造梦西游 saves were pulled out of a Flash launcher a friend built and run on the handheld, and the GBA Pokémon ROM hack runs a full team from one evolution line.
- video: https://ethanyeang.me/assets/video/handheld-rg353v-tvcast.mp4 (RG353V · Cast to the living-room TV)
- video: https://ethanyeang.me/assets/video/handheld-rg353v-zaomeng.mp4 (RG353V · 造梦西游 button mapping)

### Classic Phones / 经典设备
- zh: https://ethanyeang.me/projects/classic-phones/
- en: https://ethanyeang.me/en/projects/classic-phones/
- category: experiments; type: Nokia · Symbian · Everyday Carry
- topics: Nokia, Symbian, Retro Hardware, Everyday Carry
- en summary: A Nokia E63 and E72i back in daily use for TV, music and Bluetooth earphones — hardware from a decade ago that still does the job.
- en notes: These are the phones I wanted as a kid. The E63 has a full keyboard, the E72i is thinner, and back then watching video or playing music on either felt like the future. Second-hand units are cheap now, so I actually use them: the E63 runs a TV player and plays Star Wars on the bed, the E72i plays Eason Chan through wired earphones, and Bluetooth pairs far better than I expected.

**A design from fifteen years ago still works, still looks right.** The keys, the screen, the weight in the hand — none of that comes from a slab of glass. I like this kind of old, well-made hardware a great deal.
- video: https://ethanyeang.me/assets/video/nokia-e72i-music.mp4 (E72i · Eason Chan, 落花流水)
- video: https://ethanyeang.me/assets/video/nokia-e63-starwars.mp4 (E63 · A film before sleep)

### Flight Simulation / 飞行模拟
- zh: https://ethanyeang.me/projects/flight-simulator/
- en: https://ethanyeang.me/en/projects/flight-simulator/
- category: experiments; type: Simulator · Aviation · Collection
- topics: Flight Simulation, Aviation, COMAC, Collection
- en summary: Time in a full-scale simulator cockpit, working through the checklists — plus the carrier models and COMAC keepsakes that keep aviation within reach.
- en notes: Civil aviation has always been one of my favourite subjects. The simulator I keep going back to is a **full-scale cockpit**: overhead panel, pedestal and throttles matching the real aircraft. From cold and dark through the checklists it is the same flow as flying — only the sky outside is rendered.

A few aviation things stay within reach too: a Nimitz-class carrier model, COMAC pins and keepsakes. Not work — just a subject I am happy to spend time and money on.

### JARVIS / Reachy Mini / JARVIS / Reachy Mini
- zh: https://ethanyeang.me/projects/jarvis-reachy-mini/
- en: https://ethanyeang.me/en/projects/jarvis-reachy-mini/
- category: robotics; type: 语音助手 · 具身交互
- topics: JARVIS, Reachy Mini, XTTS, RTX Spark, RTX 5060 Ti, ComfyUI
- en summary: A local JARVIS assistant bringing XTTS speech synthesis and Reachy Mini camera and motion controls together.
- en notes: JARVIS brings local speech recognition, XTTS synthesis and Reachy Mini camera and motion controls into a visible workspace. The RTX Spark setup runs local model and ComfyUI workflows; a separate local deployment tests **XTTS voice synthesis on an RTX 5060 Ti system**. The demo shows the console with real Reachy Mini camera footage. It documents a system integration experiment, not general robot autonomy.
- video: https://ethanyeang.me/assets/video/jarvis-reachy.mp4

### Strata / Qwen3.8-Flash-Next / Strata / Qwen3.8-Flash-Next
- zh: https://ethanyeang.me/projects/strata-qwen38-local/
- en: https://ethanyeang.me/en/projects/strata-qwen38-local/
- category: tools; type: 本地模型 · RTX 5060 Ti
- topics: Qwen3.8-Flash-Next, Strata, Local Inference, RTX 5060 Ti
- en summary: A 125B Qwen3.8-Flash-Next running on one RTX 5060 Ti: 16 GB VRAM, 64K context, 74–110 tok/s decode in a real session.
- en notes: Strata serves Qwen3.8-Flash-Next from a single RTX 5060 Ti (16 GB VRAM, 47 GB RAM). The model is split into two IQ2_XS GGUF files (about 64 GB) on the SSD; once the engine is resident, 6,496 experts (8.7 GB) stay in VRAM and each request streams only about **8–17% of its experts** over **PCIe Gen5 x8** — which is how a 16 GB card keeps producing tokens.

Numbers from these real sessions: 25.4 s from cold start to ready; a short Chinese answer at 149 tokens, **75.2 tok/s**; an English one at 148 tokens, 81.8 tok/s; longer turns between 92.5 and 110.5 tok/s; prompts around 4,000 tokens prefilled at 1,091–1,449 tok/s. These are measurements from this machine at this quantization, not a guarantee for other hardware — the numbers move with the card, the quant and the context length.

### A local LLM on a Pi 5 / 树莓派 5 跑大模型
- zh: https://ethanyeang.me/projects/pi5-local-llm/
- en: https://ethanyeang.me/en/projects/pi5-local-llm/
- category: tools; type: 边缘设备 · 本地推理
- topics: Qwen2.5-0.5B, Raspberry Pi 5, LLM-8850, Edge Inference
- en summary: A Raspberry Pi 5 on Ubuntu with an LLM-8850 M.2 accelerator running AXERA's quantized Qwen2.5-0.5B: 18.15–18.42 tok/s across three runs of one prompt, and a model that introduced itself as created by Anthropic.
- en notes: The machine is a Raspberry Pi 5 on Ubuntu, with an M5Stack **LLM-8850** in the M.2 slot (M.2 M-Key 2242, AXERA open SDK). The model is AXERA's converted **Qwen2.5-0.5B-Instruct-GPTQ-Int4**, served on `127.0.0.1:8080` with every request and reply printed to the terminal.

## Measured numbers

The load log shows `AXCLWorker start with devid 0`, a run of `init N axmodel ok`, and **6454 MB** left on the card. One prompt, three consecutive runs, each closing with `hit eos, avg … token/s`: **18.42 / 18.15 / 18.37**. Those are this board, this quant and this context length — not a promise for other hardware.

## It got its own name wrong

`system_prompt` says `You are Qwen`, and the model still answered: I am Qwen, a model created by Anthropic, designed to interact with users and answer their questions. A 0.5B model does this; I left the sentence in the footage rather than edit it out.

## Its own failures

The earliest output was just a `{"token_ids": [...]}` array with the front end printing token IDs as text; that was fixed later. It also hit `save_kvcache failed` — the KV cache could not be written to a read-only directory, so the server kept running but recomputed the prefix every time.

## One question

Test question: the car wash is 50 m from home — drive or walk? The model said drive. Sent the log to a cloud model for review, and it locked onto the model's phrasing instead of the 50 m — same question, and the two models are wrong in different places.
- video: https://ethanyeang.me/assets/video/pi5-local-llm-cut.mp4 (Cut · 66 seconds)
- public data: Qwen2.5-0.5B-Instruct-GPTQ-Int4 https://huggingface.co/userresu2026/Qwen2.5-0.5B-Instruct-GPTQ-Int4

### AGIBOT G2 / Lingxi X2 / AGIBOT G2 / 灵犀 X2
- zh: https://ethanyeang.me/projects/agibot-platform-notes/
- en: https://ethanyeang.me/en/projects/agibot-platform-notes/
- category: robotics; type: 机器人平台 · 遥操作控制台
- topics: AGIBOT G2, WebXR, Teleoperation, Robot Integration, Three.js
- en summary: A local console for AGIBOT G2: a 3D skeleton and joint panel in the browser, and a headset view that pins five camera feeds in space with controller-driven arms.
- en notes: A local console I wrote for AGIBOT G2: the backend talks to the robot over the official GDK, and the front end is a single web process that the desktop and headset views share.

## Desktop: 22 joints on one screen

**22 body joints (5 waist, 3 head, 7 per arm), two 10-DOF hands, the chassis and seven camera feeds** live on one page: drag sliders, jog a joint, save and recall poses. The 3D view layers the URDF's 69 visual meshes over a joint skeleton and follows the real angles.

## Headset: camera panels pinned in space

The /xr page is where it differs: **five camera feeds (head, left eye, right eye, left hand, right hand) are panels pinned in 3D space**, each refreshing independently, so you turn your head to compare them. **Hold A or X and that arm follows the controller**; release and it stops. The trigger only closes or opens the hand; the grip button drags a panel elsewhere. A status board follows your gaze and shows whether the robot is unlocked.

## Safety is designed in

Locked by default, **auto-locking after 60 seconds idle**; a single command is capped at 0.35 rad and 0.5 rad/s and clamped again against the URDF limits. The chassis is hold-to-drive — the server sends zero velocity if it stops hearing refreshes for 0.35 s. When the robot loses power the GDK keeps returning its last frame, so the UI shows the age of that data and turns red past three seconds; otherwise stale numbers read as a live robot.

This page records platform integration and a teleoperation interface: every motion came from a person, not an autonomous policy. There is also a chaptered reel of Lingxi X2 simulation scenes, closing on the real X2 Ultra beside its simulation twin.
- video: https://ethanyeang.me/assets/video/agibot-g2.mp4
- video: https://ethanyeang.me/assets/video/g2-console-xr.mp4 (Console run · headset page and joint panel)
- video: https://ethanyeang.me/assets/video/lingxi-x2.mp4 (Lingxi X2 · Field motion)
- video: https://ethanyeang.me/assets/video/lingxi-x2-scenes.mp4 (Lingxi X2 · simulation scenes · 4:09)

### Arduino UNO Q LED Matrix / Arduino UNO Q 灯阵逆向
- zh: https://ethanyeang.me/projects/uno-q-matrix-console/
- en: https://ethanyeang.me/en/projects/uno-q-matrix-console/
- category: robotics; type: 硬件实验 · 灯阵交互
- topics: Arduino UNO Q, LED Matrix, Hardware, Interaction
- en summary: A hands-on LED matrix control experiment on Arduino UNO Q.
- en notes: The footage shows a real Arduino UNO Q LED matrix and its control interface. This is a specific hardware experiment; implementation and current support are documented in the linked repository.
- source: https://github.com/metahubaifeel/uno-q-matrix-console
- video: https://ethanyeang.me/assets/video/arduino-uno-q.mp4

### pipesay / pipesay
- zh: https://ethanyeang.me/projects/pipesay/
- en: https://ethanyeang.me/en/projects/pipesay/
- category: tools; type: 公开仓库
- topics: Python
- en summary: Linux live dictation for Wayland — speak, see text, auto-copy. MIT.
- en notes: See the public repository for implementation notes, setup instructions and current status. Linux live dictation for Wayland — speak, see text, auto-copy. MIT.
- source: https://github.com/metahubaifeel/pipesay

## Public Hugging Face repositories
- dataset: so101-20260422_0337 — https://huggingface.co/datasets/userresu2026/so101-20260422_0337
- dataset: so101-20260422_0445 — https://huggingface.co/datasets/userresu2026/so101-20260422_0445
- dataset: b601-dm-20260718_0031 — https://huggingface.co/datasets/userresu2026/b601-dm-20260718_0031
- dataset: b601-dm-20260718_0053 — https://huggingface.co/datasets/userresu2026/b601-dm-20260718_0053
- dataset: b601-dm-icehockey-smoke — https://huggingface.co/datasets/userresu2026/b601-dm-icehockey-smoke
- dataset: b601-dm-icehockey-v1 — https://huggingface.co/datasets/userresu2026/b601-dm-icehockey-v1
- dataset: lekiwi-bubble-hitting-clean9 — https://huggingface.co/datasets/userresu2026/lekiwi-bubble-hitting-clean9
- dataset: lekiwi-bubble-hitting — https://huggingface.co/datasets/userresu2026/lekiwi-bubble-hitting
- dataset: b601-lingbot-vla2 — https://huggingface.co/datasets/userresu2026/b601-lingbot-vla2
- model: Qwen2.5-0.5B-Instruct-GPTQ-Int4 — https://huggingface.co/userresu2026/Qwen2.5-0.5B-Instruct-GPTQ-Int4

## Scope note
Project pages document specific experiments and their recorded environment. They are not general performance guarantees. Forks remain attributed to their upstream authors. Public datasets and model cards are maintained on Hugging Face.