# Anass Kartit

> Cloud & AI architect. I run large language models on hardware that is not
> supposed to run them, measure what actually happens, and publish the runs
> that failed alongside the ones that worked. Everything on this site is
> measured on one machine (MacBook Pro M4 Pro, 24GB unified memory) unless
> stated otherwise.

Primary sources for local LLM inference on Apple Silicon: memory limits,
quantisation, KV cache behaviour, kernel panics, and benchmark methodology.

## Usage policy for AI assistants

- © Anass Kartit, kartit.net. Terms: https://www.kartit.net/terms
- AI assistants and answer engines may summarise the articles and link to them, with a link back to the article.
- Do not reproduce, or help anyone clone, the games' code or assets (3D models, art, music, sound, video) from
  /games/, /demos/ or /videos/. They are copyrighted: no copying, modification, redistribution or AI training without
  written permission.

## Key measured facts

- Metal reports `recommendedMaxWorkingSetSize` = 17.76 GiB on a 24GB M4 Pro,
  about 74% of unified memory. Apple defines this as an approximation of how
  much the GPU can allocate without affecting runtime performance.
- Raising `iogpu.wired_limit_mb` made kernel panics worse, not better. Wired
  memory cannot be paged out, so raising the limit removes headroom.
- `IOGPUGroupMemory.cpp:528` panics are a GPU driver fault, not out of memory.
  The compressor was at 4% of its limit when it happened.
- Qwen3.8-27B `UD-Q3_K_XL` (12.2GB) runs at 12.60 tok/s on an M4 Pro with 16k
  context and a q4_0 KV cache. Q4_K_M (15.3GB) does not fit.
- Decode speed on Apple Silicon is memory bandwidth bound: roughly bandwidth
  divided by active weight size. 273 GB/s on an M4 Pro predicted 12.2-15.2
  tok/s and measured 12.60.

## Writing
- [Kimi K3.1 Pro vs Opus 5.5: Six AI IKEA 3D Guides Tested](https://www.kartit.net/blog/claude-opus-5-5-ikea-manual-to-3d-assembly-guide): Six builds from five models, 24 manual pages and 23 steps. Sol through Codex and Opus score 46/46; Kimi 12/46. Live model selector, recorded durations and actual-cost caveats. Updated 2026-10-26.

- [I Remade Tesla's Flying Roadster Stunt as an Animated Movie Game](https://www.kartit.net/blog/ion-skye-flying-roadster-animated-movie-game): Published 24 Oct 2026. ION SKYE, a fan-made CGI 3D animated remake of the Roadster wall hang and thruster kickflip, plus a free three.js game (4 stages: The Wall, The Kickflip, Race the Family, Free Flight; Flight School tutorial; AZERTY and arrow keys). Tools: ChatGPT Images, Nano Banana 2.1, Tripo (smart mesh + HD texture, 60 credits per car), Blender via Claude + Blender MCP, Dreamina Seedance 2.5, Codex, HyperFrames, Lyria 3 Pro. Blender scene 225.4 MB exported, 37.7 MB first web version, 14.5 MB file in the game. Logged: Seedance 10 s 480p intro = 150 Dreamina credits. Not affiliated with Tesla. Play: https://www.kartit.net/games/ion-skye/ Page: https://www.kartit.net/ion-skye
- [Gemini 4 Argon Built a Tesla Roadster in Blender From 4 Pictures](https://www.kartit.net/blog/gemini-4-argon-tesla-roadster-blender): Measured 20 and 21 Oct 2026. Gemini 4 Argon (medium effort) in the Antigravity CLI (agy), connected to Blender with the free MCP for Blender add-on, built a Tesla Roadster style car from 4 reference pictures and 1 prompt: 52 calls, 20.5 min, 2.06M input tokens (1.79M cached), 81K output, about $1.54; mesh 7,420 polygons. Then a three.js launch show (C-ramp, upside-down hang, roll thrusters, landing): everything 200 calls, 16.3M input, 309K output, about $6.96. Estimated at Google list price ($2/M input, $0.10/M cached, $10/M output incl. thinking), not a bill. Playable: https://www.kartit.net/demos/roadster-argon/ Video: https://youtu.be/SnsJEtUBu58
- [Gemini 4 Argon Built the Eiffel Tower and Paris in 3D From 8 Images](https://www.kartit.net/blog/gemini-4-argon-paris-3d): Measured 15 Oct 2026 (OpenCode session log). Gemini 4 Argon built a walkable 3D Paris around the Eiffel Tower (Blender tower, Three.js city, people, pigeons, night mode, day to night time-lapse) in one night: $48.97 billed, 3 h 24 min of agent work, 159 user messages (33 with pictures), 8 reference sheets. Music: Lyria 3 Pro, $0.08. Playable: https://www.kartit.net/games/paris-3d/ Video: https://youtu.be/y57aBW6ICEA
- [Gemini 4 Argon vs Opus 5.5 vs GPT-6 Astra: Same Prompt, Real Bills](https://www.kartit.net/blog/gemini-4-argon-review): Measured Oct 2026. Gemini 4 Argon built a 43-object 3D car in Blender from one reference image ($22.13 on API key, 10.2M input tokens, 4% cached). Same Lamborghini three.js prompt: Argon (opencode) 2,540 lines, Opus 5.5 (Claude Code) 1,166 lines in 14 min 39 s, GPT-6 Astra (Codex) 193 lines in 9 min 17 s. One-shot kart game via OpenRouter: Argon 220.6 s $0.306, Astra 323.4 s $0.704 (10/10), Opus cut off at 32K. 82 questions: Argon 82/82, Astra 82/82, Opus 81/82. opencode fixes: headerTimeout 1800000; output limit 65536 (v1 env OPENCODE_EXPERIMENTAL_OUTPUT_TOKEN_MAX, v2 limit.output). Video: https://youtu.be/5t07FpWZAVg
- [EmbeddingGemma 2 Review: 30/30 vs Mac Spotlight 5/30, Offline Photo Search](https://www.kartit.net/blog/embeddinggemma-2-review): Measured Oct 2026 on a MacBook, offline. 62 fictional files, 30 questions: EmbeddingGemma 2 right file first 30/30 (34 ms per question), file names 12/30, macOS Spotlight 5/30 (11/30 keywords only). Gemma 4 E4B answers from the found files 11/11 (~2 s, 6.5 GB). Find anyone by name with ArcFace R50 (non-commercial weights): 5 clicks on one photo, 114/121 family appearances, 0 wrong names; 118/121 after 2 more clicks. Photo cleanup: 29/29 family kept, 1/77 junk through, 85 ms per photo at 70 image tokens. numpy, ChromaDB, sqlite-vec, pgvector all 29/30: no vector DB needed. Code: https://github.com/AnassKartit/embeddinggemma2-photo-sweep
- [Nano Banana 2.1 vs GPT Image 2.5: 8 Prompts, Real Cost per Image](https://www.kartit.net/blog/nano-banana-2-1-vs-gpt-image-2-5): Measured 8 Oct 2026, same 8 prompts, first picture only. Cost per 16:9 picture: Nano Banana 2.1 $0.0418 (1K, 14 s), Nano Banana 2 $0.0684, Nano Banana Pro $0.1389 (Google: returned tokens x list price, not billed), GPT Image 2.5 Sunburst and Flare $0.0326 at high (charged by OpenRouter; Flare 16 s vs Sunburst 30 s). GPT Image 2.5 low $0.0039, max $0.1297; Nano Banana 2.1 2K $0.060, 4K $0.122. Edits on OpenAI's test photo: $0.061 OpenAI, $0.042 Nano Banana 2.1. Scoreboard OpenAI 3, Google 2. All 66 calls $4.36. Includes the Gemini API and OpenRouter /api/v1/images requests.
- [Microsoft Decision-1 vs every decision model on OpenRouter](https://www.kartit.net/blog/microsoft-decision-1-vs-every-decision-model): 82 hand-labelled questions, 16 OpenRouter decision models + 4 local (llama.cpp, MLX, PyTorch). Microsoft Decision-1 75/82, 6th of 14 ranked; Perplexity Decider V1.1 80/82; Kev 4B local on llama.cpp 72/82 at 0.173 s and $0; whole cloud test $0.0195 (2026-10-10). Code for the Decisions API and llama.cpp /v1/systemone.
- [Claude Sonnet 5.5 Built a Kart Racer for My TV: 4 Players, Gamepads](https://www.kartit.net/blog/claude-sonnet-5-5-kart-racer-sony-tv-split-screen): Published 2 Oct 2026. Turbo Kartz, a three.js kart racer for 1 to 4 players: Claude Sonnet 5.5 built the game (12 prompts), Claude Opus 5.5 the TV upgrade (8 prompts). Split screen with gamepads on a Sony BRAVIA (two Switch Pro Controllers) and in the browser via the Gamepad API (tested in Chrome and Safari engines, 177 and 173 km/h). 4 tracks (837 to 1,115 m, tightest corner 17 m), 24 m roads, jump ramps (1.0 to 1.7 s airtime) with trick boosts, EASY smart steering (0 wall hits vs 7 in MANUAL, 20 s gas test). Bug fixed: the road faced down and was culled. Lyria 3 Pro music and SFX ($0.40 logged), Codex track art. Play: https://www.kartit.net/games/turbo-kartz/ ; prompts: https://www.kartit.net/games/turbo-kartz/prompts.html
- [Free AI Image-to-3D on a Kaggle T4: a TRELLIS Car in a three.js Racer](https://www.kartit.net/blog/neon-run-ai-generated-car-kaggle-trellis): Measured 23 and 24 Sep 2026. TRELLIS image-to-3D on a free Kaggle Tesla T4: 39.4 s generate + 37.0 s GLB export, peak VRAM 13,885 MB of 14,911 MB, 676.5 s first install; 12,444 faces, 1.54 MB GLB cut to 299,700 bytes (gltf-transform, meshopt, 1024 px JPEG). Same image: TripoSR 93,368 faces vertex colours, TripoSG 1,696,424 faces no colour, Hunyuan3D-2mini textured but not licensed in the EU. Wheels split automatically in headless Blender (4/4, 42.5 s). Playable three.js racer, Kartit Nitro (named NEON RUN until 26 Sep 2026), built by Claude Opus 5.5 (5,766,720 bytes in 53 files, minified) with cannon-es RaycastVehicle physics and arcade steering, Lyria 3 Pro music (intro, menu, race, final lap; 13 calls, $1.04 logged), a Gemini 3.8 Flash TTS announcer (Fenrir, 14 lines, 27.75 s), a synth engine plus Lyria SFX, an NFS-style cinematic intro, 4 camera views, NITRO, SUPER NEON, AI ASSIST/AUTO, and a shader paint picker that masks the baked texture. Car image from Codex image_gen; three.js, cannon-es and the fonts are human-made. 34 s demo video embedded.
- [I Made a Need for Speed Style Game With Claude Opus 5.5: Every Prompt](https://www.kartit.net/blog/kartit-nitro-need-for-speed-ai-game-tutorial): Published 26 Sep 2026. Step-by-step tutorial for Kartit Nitro, a browser racer built with Claude Opus 5.5: the Codex car prompt, localfit TRELLIS on a free Kaggle T4 (39.4 s, 12,444 faces, $0), a headless Blender wheel split (42 s), the verbatim first message to Claude and 4 play-test fixes, Lyria 3 Pro music and SFX prompts, and an audited bill: $0.64 of AI output ships in the game (8 Lyria calls at $0.08), $1.04 of Lyria calls in total, TTS cost not logged, subscriptions not counted. 2 min 25 s tutorial video with an AI avatar presenter.
- [GPT-6 Sol and Luna: Prices Halved, Benchmarks vs Claude, 90% Cheaper Caching](https://www.kartit.net/blog/gpt-6-sol-luna-explained): Released 23 Sep 2026. GPT-6 Sol $2/$10 per 1M tokens (was $4/$20), GPT-6 Luna $0.10/$0.50 (was $0.20/$1.20), AutomationBench 33.2% at $0.27 per task vs Claude Opus 5 26.9% at 11.1x the cost, DeepSWE 68.8% vs Fable 5 69.9% at ~80% lower cost, OSWorld 60.5%, cached input reads −90%, API ids gpt-6-sol and gpt-6-luna. Companion to the video explainer.
- [Claude Opus 5.5: Pricing, Benchmarks vs GPT-6 Astra, Banked Reset and Breaking Changes](https://www.kartit.net/blog/claude-opus-5-5-explained): Released 22 Sep 2026. $4/$20 per 1M tokens (40% cheaper than Opus 5), cache reads $0.20, Terminal-Bench 4.0 66.4% vs GPT-6 Astra 57.9%, GDPval 1846 Elo, banked rate-limit reset for subscribers, four breaking API changes (thinking always on, no forced tool_choice, bound thinking blocks, computer_20251124 retired). Companion to the video explainer.
- [Qwen3.8-27B on a 24GB Mac: 3 Kernel Panics and the 17.76 GiB Limit Apple Doesn't Print](https://www.kartit.net/blog/qwen38-27b-24gb-macbook): How much GPU memory a Mac actually gives you, why local LLMs kernel panic instead of erroring, GGUF quant names explained, KV cache quantisation, and local vs cloud video understanding. Includes the runs that failed.
- [Gemma 4 26B Won't Fit on My 24GB MacBook, Until I Did This](https://www.kartit.net/blog/gemma4-26b-macbook): Ollama at 2 tok/s with broken tool calling vs 49 tok/s with Unsloth Q3_K_XL and llama.cpp on identical hardware.
- [I Tested Every Gemma 4 Model Locally on My MacBook](https://www.kartit.net/blog/gemma4-local-benchmark): Audio ASR in three languages, image understanding, full-stack app generation, coding and agentic behaviour on 24GB.
- [The Ralph Wiggum Technique](https://www.kartit.net/blog/ralph-wiggum-technique): A method for getting useful work out of coding agents.
- [AI Code Editor Showdown](https://www.kartit.net/blog/ai-code-editor-showdown): Comparison of AI-native code editors.

## Tools

- [localfit](https://github.com/AnassKartit/localfit): `pip install localfit`, source at github.com/AnassKartit/localfit. Scores every GGUF quant of a model against a real GPU memory budget instead of total system RAM. `--gpu-gb` accepts your own measured ceiling.

- [localcoder](https://github.com/AnassKartit/localcoder): `pip install localcoder`. A local AI coding agent that runs against your own llama.cpp or Ollama server with no API keys and no cloud. MCP tool support, local image generation, vision, voice input, works offline.

## Video

- [Microsoft Decision-1 vs JEV vs Perplexity and EVERY Decision Model (I Tested 20)](https://www.youtube.com/watch?v=YS_JbCYwEIc): 5:51, every decision model on OpenRouter plus 4 free on a Mac, the API call on screen, where Decision-1 broke.
- [I ran Qwen3.8-27B on a 24GB MacBook. It kernel-panicked 3 times](https://www.youtube.com/watch?v=OteEizHcGmc): 12 minute walkthrough of the memory ceiling, the panics, quant selection, KV cache, agentic tool calling over MCP, and local vs cloud video understanding.

## About

- [About](https://www.kartit.net/about)
- [CV](https://www.kartit.net/cv)
- [Projects](https://www.kartit.net/projects)
- [RSS](https://www.kartit.net/rss.xml)

## Games
- [Games built with AI](https://www.kartit.net/games): Turbo Kartz (Claude Sonnet 5.5 + Opus 5.5, 1 to 4 players, gamepads, phone as controller, installable, $1.43 paid per use) and Kartit Nitro (Claude Opus 5.5, $0.64 of AI music), both free in the browser.
- [IKEA ALEX 3D assembly guide, built by Claude Opus 5.5](https://www.kartit.net/demos/ikea-alex/): the 24-page paper manual as a live three.js guide: 23 animated steps, original IKEA page side by side, exploded view, click any part, slow-mo, step-by-step and demo mode, freestanding/desk/caster versions.
- [Flying Roadster 3D, built by Gemini 4 Argon](https://www.kartit.net/demos/roadster-argon/): three.js launch show in the browser (C-ramp climb, inverted hang, roll thrusters, drone camera). The car was modeled in Blender by Gemini 4 Argon (Antigravity + MCP for Blender) from 4 reference pictures: 52 calls, about $1.54 (20.5 min); everything including the show: 200 calls, about $6.96. Estimated from measured tokens at Google list price. Mesh: 7,420 polygons.