Skip to content

About

Personal LLM experiments: original Markdown reports, measured results, failures, and a searchable static archive

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

LLM Test Reports

English | 简体中文

My LLM test archive: models and inference engines observed through real tasks, failures and human feedback. English is the default; full Chinese editions use _CN. Reports, indexes and the website are bilingual.

Browse the report site · Full Markdown reports · Add a report

The archive contains 54 experiment, parameter or comparison records and 77 reports, each in full English and Chinese editions. Multiple review rounds belong to one test chain; different models can share a report. These counts are not independent sample counts.

Workstation and system

Component Current verified configuration
Computer MacBook Pro Mac17,6 · Apple M5 Max
CPU / GPU 18-core CPU / 40-core GPU
Unified memory 128 GB system label; 137438953472 bytes (128 GiB)
System macOS 27.0.1 · build 26A434 · arm64 · Metal 4
Snapshot date 2026-10-20 22:03:27 UTC

This is a current reference, not an assumed environment for every historical test. Each report identifies whether test-time OS evidence survives; LAN PC and cloud tests distinguish client from inference host. The separate Laya M3 Max runtime benchmark retains its reported environment, and the service evaluation retains its Darwin 27.0.0 kernel version. See full configuration and historical limits.

Artifacts

Artifact Separate repository Play
Neon Bastion · mlx-dspark / Qwen3.8-27B 8bit / Pi llm-test-2026-09-01-game-neon-bastion GitHub Pages
Tank Battle · MLTF / Qwen3.8-27B 8bit q8c / Pi llm-test-2026-10-20-game-tank-battle GitHub Pages

Artifact repositories use llm-test-YYYY-MM-DD-{type}-{name}, dated by test start. Neon was renamed with Git history preserved; Tank Battle is the final four-round snapshot. Other Neon model versions have their own reports and are distinct implementations.

Experiment index

Date Experiment / parameters Model and runtime Report
2026-10-21 Splash affine8 native 256K context stress test Qwen3.8-27B MLX affine8 / INT8 KV PERF-11
2026-10-21 Pi capability exam: Qwen3.8-27B 8-bit on MLTF / Splash Qwen3.8-27B 8-bit (MLTF q8c / MLX affine8) CAP-01
2026-10-21 Qwen3.8-27B: three local 8-bit configurations and GGUF record Qwen3.8-27B MLX affine8 / MLTF q8c / GGUF Q8_0 PERF-10
2026-10-21 Splash vs MLTF / PERF-06 128K vision replay Qwen3.8-27B MLX affine 8bit / MLTF q8c; INT8 KV PERF-09
2026-10-20 MLTF × Pi / four rounds of 3D Tank Battle MLTF 0.1.2 + Qwen3.8-27B 8bit q8c + Pi TANK-01
2026-10-19 MLTF vs mlx-vlm / 128-request quality comparison Qwen3.8-27B 8bit; MLTF vs mlx-vlm PERF-08
2026-10-19 MLTF / real Qwen, Pi and Hermes CLI checks Qwen3.8-27B 8bit; MLTF PERF-07
2026-10-19 MLTF vs mlx-vlm / 128K vision performance Qwen3.8-27B 8bit; MLTF vs mlx-vlm PERF-06
2026-10-19 mlx-vlm / vision and DFlash2 stability ladder Qwen3.8-27B 8bit; mlx-vlm vision + DFlash2 stepped stability PERF-05
2026-10-19 mlx-dspark / 128K and 256K long context Qwen3.8-27B 8bit; mlx-dspark 128K PERF-04
2026-10-11 AI Werewolf / ten DeepSeek V4.1 Flash games DeepSeek V4.1 Flash API; cloud CLOUD-01
2026-10-06 Repair Café / Pi coding and approximately 151 minutes of soak Pi → LAN PC Qwen3.8 Flash-Next; code stored on Mac PC-02
2026-10-06 Qwen3.8 Flash-Next / initial 21-request API check LAN PC Qwen3.8 Flash-Next service; Mac client PC-01
2026-09-21 Local Laya service / 72 questions and concurrency limits aac6fef/laya-multilingual-mlx; GPU FP16; 1024 tokens LAYA-01
2026-09-21 Jev vs local Laya / the same 72 questions Jev cloud proxy compared with local Laya JEV-01
2026-09-19—09-20 MTPLX 27B Quality / cold prefill and near-128K soak MTPLX; Qwen3.8-27B Optimized Quality PERF-02
2026-09-19—09-20 MTPLX Flash-Next / cold prefill and near-128K soak MTPLX; Flash-Next Optimized Speed PERF-01
2026-09-08—09-16 Neon Bastion / ds-v4-flash-vision 55-task run ds-v4-flash-vision + DeepSeek Harness; cloud comparison NEON-03
2026-09-07—09-16 Neon Bastion / MTPLX 55-task rerun MTPLX Qwen3.8-27B Optimized Speed + Pi NEON-02
2026-09-05—09-06 Tide Isle / R1, R2 and PLAY copy MTPLX Qwen3.8-27B Optimized Speed + Pi TIDE-01
2026-09-01—09-04 Neon Bastion / seven 8bit rounds mlx-dspark + Qwen3.8-27B-8bit + DFlash2 + Pi NEON-01
2026-08-31—09-01 (directory / conversation) Neon Bastion / Flash-Next version Directory label: MTPLX Flash + Pi FLASH-04
2026-08-31—09-01 Gare Saint-Lazare MTPLX qwen3.8-flash-next-optimized-speed + Pi FLASH-03
2026-08-31—09-01 Marble Workshop MTPLX qwen3.8-flash-next-optimized-speed + Pi FLASH-02
2026-08-31—09-01 Orbital Garden MTPLX qwen3.8-flash-next-optimized-speed + Pi FLASH-01
2026-08-25—08-26 JSON Lens / Qwen Qwen3.8-27B JSON-02
2026-08-25—08-26 JSON Lens / Ornith Ornith 1.5 35B-A3B JSON-01
2026-08-23—08-24 Frog Choir / initial and repair rounds MTPLX Optimized Quality + Pi / Qwen Code FROG-01
2026-08-22 oMLX RC / oQ4e-mtp dashboard observation oMLX 0.6.3rc2 + Qwen3.8-27B-oQ4e-mtp PERF-03
2026-08-22—08-23 LLM memory-bandwidth teaching page / three review rounds oMLX RC Qwen3.8-27B-oQ4e-mtp + Qwen Code EARLY-04
2026-08-22 Paris 3D oMLX Qwen3.8-27B + Qwen Code EARLY-03
2026-08-21 Arcology control room / Pi attempts and Qwen Code delivery oMLX True2456/Qwen3.8-27B-AWQ-5.0bpw; Pi / Qwen Code coding-reason EARLY-02
2026-08-21 approximately (conversation) Boston itinerary map / Pi Pi + oMLX Qwen3.8-27B AWQ 5bpw EARLY-01
2026-06-02—06-04 Tetris / Qwen3.5-122B-A10B-4bit Claude Code + oMLX; Qwen3.5-122B-A10B-4bit TET-04
2026-06-02—06-04 Tetris / Qwen3.6-35B-A3B-MLX-8bit Claude Code + oMLX; Qwen3.6-35B-A3B-MLX-8bit TET-03
2026-06-02—06-04 Tetris / Qwen3.6-27B-UD-MLX-4bit Claude Code + oMLX; Qwen3.6-27B-UD-MLX-4bit TET-02
2026-06-02—06-04 Tetris / gemma-4-31b-it-8bit Claude Code + oMLX; gemma-4-31b-it-8bit TET-01
2026-06 (original report) Shopping MVP / gemma-4-31b-it-8bit Claude Code + oMLX; gemma-4-31b-it-8bit SHOP-04
2026-06 (original report) Shopping MVP / Qwen3.6-27B-UD-MLX-4bit Claude Code + oMLX; Qwen3.6-27B-UD-MLX-4bit SHOP-03
2026-06 (original report) Shopping MVP / Qwen3.6-35B-A3B-MLX-8bit Claude Code + oMLX; Qwen3.6-35B-A3B-MLX-8bit SHOP-02
2026-06 (original report) Shopping MVP / Qwen3.5-122B-A10B-4bit Claude Code + oMLX; Qwen3.5-122B-A10B-4bit SHOP-01
2026-05-27 Cafe POS from zero / Config E Hermes + oMLX Qwen3.6-35B-A3B-MLX-8bit HERMES-06
2026-05-27 Cafe POS from zero / Config D Hermes + oMLX Qwen3.6-35B-A3B-MLX-8bit HERMES-05
2026-05-27 Cafe POS from zero / Config C Hermes + oMLX Qwen3.6-35B-A3B-MLX-8bit HERMES-04
2026-05-27 Cafe POS from zero / Config B Hermes + oMLX Qwen3.6-35B-A3B-MLX-8bit HERMES-03
2026-05-27 Cafe POS from zero / Config A Hermes + oMLX Qwen3.6-35B-A3B-MLX-8bit HERMES-02
2026-05-27 Hermes POS gauntlet / Cart & Pricing Hermes + oMLX Qwen3.6-35B-A3B-MLX-8bit; Config A HERMES-01
2026-05-26 3D Solar System / interrupted 40B run Qwen3.6-40B-Claude-4.6-Opus-Deckard-Heretic-Uncensored-Thinking-8bit SOL-06
2026-05-26 3D Solar System / gemma-4-31b-it-8bit Hermes + oMLX; gemma-4-31b-it-8bit SOL-05
2026-05-26 3D Solar System / Qwen3-Coder-Next-8bit Hermes + oMLX; Qwen3-Coder-Next-8bit SOL-04
2026-05-26 3D Solar System / Qwen3.6-27B-MTPLX-Optimized-Speed Hermes + oMLX; Qwen3.6-27B-MTPLX-Optimized-Speed SOL-03
2026-05-26 3D Solar System / gemma-4-26b-a4b-it-bf16 Hermes + oMLX; gemma-4-26b-a4b-it-bf16 SOL-02
2026-05-26 3D Solar System / Qwen3.6-35B-A3B-MLX-8bit Hermes + oMLX; Qwen3.6-35B-A3B-MLX-8bit SOL-01
Date unverified Four model self-introduction HTML specimens Filenames: 4bit / 8bit / fp16 / flash; full configuration unverified SELF-01

Contents

English files have paired _CN Chinese editions. Technical code and screenshots retain their source-language evidence.

reports/originals/      # full reports
reports/experiments/    # entries grouped by experiment
assets/                # original figures and screenshots
data/                  # paired catalogs, environments, provenance and translation audit
site/                  # plain HTML / CSS / JS
scripts/               # registration, build and validation
templates/             # paired report templates
.github/workflows/pages.yml

Local preview

npm ci --ignore-scripts
npm run build
npm run check
npm run preview

Open http://127.0.0.1:8437/ for English or http://127.0.0.1:8437/index_CN.html for Chinese. dist/ contains static files without a server API; GitHub Pages uses the same build.

Evidence and limits

Original measurement definitions, failures, independent model implementations and human interventions are retained. Missing source is not fabricated, and runtime observations are not presented as controlled benchmarks. This organization, translation and publication did not repeat the historical generation experiments.

Publication relocates local links, links separate artifacts or marks unpublished historical files; private account paths and LAN addresses are generalized. Full translations preserve code, commands, numeric literals and original images. Translation is not a new performance measurement. Source, initial import and current localized hashes are in provenance.json; source-code preservation checks are in translation-audit.json. Full private sessions and client configuration were not imported.

Reports and the website are maintained here; artifact code and licenses belong to their separate repositories.

About

Personal LLM experiments: original Markdown reports, measured results, failures, and a searchable static archive

Topics

Resources

Contributing

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages