Index of published work

Public summary for readers outside this repo: five servers from 8K to 1M, agent turn time, prompt cache reuse, memory, a quality check and the commands. mlx-serve: 8.8 s per agent turn at 128K and 13.0 s at 512K. Splash 1.4.0 with Swift 1.5: 6.7 s at 512K and 9.3 s at 1M, but a change halfway re-prefills the whole prompt.

Public summary for readers outside this repo: what happens to speed and waiting time when two, three or four agent sessions share one local server, and how to set up the launcher for it. 2 sessions is the maximum; each runs about 2× slower; running two together saves no total time.

Four runtimes: headline numbers, bars per metric, gates per band, test coverage, a cheap quality check (HumanEval, tool calls) and the September uncensored stage. n2 wins with T_turn 8.23 s at 32K and 8.84 s at 128K.

October 2026 stack: mlx-serve 26.10.1 and 26.10.3, oMLX 0.7.0, MTPLX 2.12.2, ds4 and Splash 1.4.0. Lines per context for T_turn, warm TTFT, cold TTFT, decode, prefill and wired memory. Splash has the lowest T_turn in every band and is the only runtime at 1M; mlx-serve leads the four runtimes of the engine round; MTPLX refuses 256K (planned fit 237,568 tokens).

The effort line first or last in the prompt, on Swift 1.5 and iQ-MLX: effort switch at 128K, prefix reuse on short prompts, obedience per level, time and tokens per question. Last: the switch drops from 167 s to 0.8–5.8 s and accuracy holds, but short prompts lose the shared prefix (53/55 → 0/55 reused).

Two 4.7-bit MLX packs of Qwen3.8-Flash-Next on mlx-serve 26.10.1: time per question as dot plots, tokens and generation speed, accuracy on six suites. Same accuracy; with the recommended sampling, 2–3× less time to the final answer on GPQA and LiveCodeBench, because Swift 1.5 thinks with fewer tokens.

The campaign's first dashboard: verdict in prose, scoreboard per band, T_turn and TTFT lines, cache hit per scenario, wired memory and the 512K probe.

Campaign dashboard: sustained decode per arm, runtimes and quants, tests, gate verdicts, queue and coverage matrix.

Five configurations that reached the native context: decode, prefill, warm and cold effective throughput, total wait, peak RAM and cache hit per scenario.

Splash 1.3.1 + DFlash2 is the fastest of three servers: 17.7 s (thinking effort xhigh) / 15.9 s (medium) per agent turn at 128K, 0.69× the mlx-dspark baseline, still about 2× Qwen3.8-Flash-Next (8.6 s). mlx-serve 26.10.2 was slower than the baseline. Funnel, run command, limitations.

One-page summary of the 9 datasets: decode by model size, long-context wait, quality, hallucination and RAG, reasoning probes, cross-harness consistency and SVG quality. It was the site's landing page until 2026-09-14.

Accuracy (n=100), LiveCodeBench (n=50), tool calling, throughput, speed probe, prefill and elapsed time of the local models on LM Studio / MLX. Filters by model, tier, architecture and quant.

Published scores (MMLU-Pro, GPQA, SWE-V, AIME, LCB v6, T-Bench, MMMU) of the local models against frontier and open-weight reference models.

Ranking of 9 local models and an 89-task × 9-model matrix: who solved what, by category, difficulty and time budget.

84 SVGs from 6 harness × model pairs, graded requirement by requirement by a single judge: leaderboard, head-to-head, per-question detail and a filterable gallery.

28 SVGs from Codex × gpt-5.6-sol on seven SVGBench prompts, four takes each (v1, v2, v3 and animated). Unscored: a gallery to compare the takes side by side.

Five agent stacks build the same 3D game from the same brief in one session: time, output tokens, lines shipped, bundle size and defects found.

One local model (Qwen3.8-Flash-Next), three design skills and a no-skill baseline build five landing-page iterations: time, output tokens and the pages side by side.

Performance observed in day-to-day opencode use, per model and provider: TPS, TTFT, prefill, TPS per context, context size, cache hit and cost. Field usage, not a controlled benchmark.

Qwen3.8-Flash-Next's work next to frontier models' work on the same tasks: graded SVG drawings, the built games and the landing pages, each with its numbers. 112 drawings from four newer pairs (gpt-5.6-sol, gpt-6-astra and two DeepSeek models) have no verdict yet and are marked unscored.

The Agent Build-Off 3D game built by opencode with local Qwen3.8.

The same stack with the superpowers skills enabled.

The 3D game built by Claude Code running local Qwen3.8.

The 3D game built by Claude Code with hosted Claude Opus 5.

The 3D game built by Codex with hosted gpt-6-astra (multi-model).

Five Layout Skill Bench landing pages built with the design skill, with a switcher between them.

Five Layout Skill Bench landing pages built with the taste skill.

Five Layout Skill Bench landing pages built with the second version of the taste skill.

Five Layout Skill Bench landing pages from the baseline run, built without a design skill.