Public copy: local paths, private network addresses and personal identifiers are generalized in prompts, handovers and downloads. The original design requests were the same for each model. CAD geometry is unchanged.
Methodology
DeepSeek harness adjustment: After repeated reasoning-only replies exhausted 32,768 tokens, increasing the reply allowance did not provide a timely return to tools. The two DeepSeek runs were resumed with a 2,048-token reasoning budget, enforced by TensorFold through a run-local transport. The existing high effort, temperature 1.0 and top-p 0.95 remain unchanged. This closes the reasoning block and permits an answer or tool call; it is not a wall-clock deadline. The gateway repetition guard only evaluates completed replies. These are adjusted-harness results, not untouched default-runtime results. The original design prompts remain unchanged.
Two design tasks, fifteen requested models, thirty independent runs. Mistral Large 4, GLM 5.3, Kimi K3, Gemini 4 Argon, MiniMax M3.1 Flash, GLM 5.3 Flash and GPT-OSS 120B were added later, using OpenCode. The two GPT-OSS runs restrict OpenRouter routing to Cerebras only; other runs use their ordinary provider routing with its default model settings, a separate sandbox and the same verbatim prompts and font. Every model starts with exactly the same original prompt for a given task, including the working conditions below. Each run starts with its own top-level directory and a local copy of FillaPrint Regular and its license. The font and prompt hashes are checked for equality across runs.
Native Claude Code, Codex and OpenCode TUIs are operated through tmux. Each model uses its default effort; no effort variants are compared. A filesystem sandbox hides other run directories, previous designs and session histories, and exposes the run’s own files plus read-only existing CAD runtimes. Research is allowed through the specified proxy. There are no installs, physical prints, slicing, design passbacks or shared agent work.
The twelve original cloud entries and fourteen later entries (Mistral-hosted Mistral and GLM; OpenRouter-hosted Kimi, Gemini, MiniMax, GLM Flash and GPT-OSS 120B) each received one design request with no subsequent design feedback. The model could research, build, measure and fix its own errors within that turn. After the four Spark runs had continued without delivering exported parts, the owner requested the same finish-focused follow-up for all four. Its exact text appears beneath each original prompt and in the follow-up download. Those four results are marked as receiving follow-ups and are not strict one-shot submissions. They also received a second identical request to wrap up. The two remaining Qwen runs subsequently received the same instruction to finish within 20 more tool calls. The exact texts appear above each thumbnail grid. The follow-up supplied no dimensions, geometry corrections or design solution. Setup attempts interrupted while establishing the launch environment are excluded and archived privately; the published entries begin with the clean native-TUI runs. Some native sessions were resumed after restoring status-line configuration, with only “Continue the original request.” supplied; no design feedback or correction was given. Qwen initially exhausted its native 16,384-token reply cap with reasoning-only responses before tool execution. DeepSeek later exhausted its 32,768-token reply cap during reasoning as well. Both tasks for each Spark model were resumed with a 65,536-token reply allowance, preserving their default thinking settings and original requests. These runtime interventions are part of the reported conditions. Spark sessions use the Qwen tasks route for automatic tool-permission classification. Some Qwen tool calls were delayed by classifier timeouts and later retried by the model; those failed classifications are infrastructure delays, not CAD execution or design failures. Those infrastructure interruptions make raw wall-clock timings unsuitable for a strict speed ranking. Existing RAM experiments and the previous overnight trials are excluded from this comparison.
Review examines the delivered STEP/STL geometry independently. For the mailer: real supported module and bag sizes, closure engagement and travel, protection of contacts, removal access, compactness and readable labels. For Gridfinity: base compatibility, usable compartment capacity, hardware separation, finger access, labels, lifting and stacking, footprint and extension to further sizes. A fatal functional problem prevents acceptance even when a mesh is watertight.
GLM 5.3 transport failures: Two DIMM responses ended with recorded provider.transport connection-reset errors after 48.39 and 16.76 minutes, each with retry attempt 2 and no token counters. The event metadata carries status 200 alongside these failures. The records do not locate the reset at the provider, client or intervening network. These failed/retried calls confound wall-clock speed comparisons; their missing usage is not treated as zero consumption.
Timing, where reported, will distinguish setup, active run and finished handover. A single run per model is a qualitative comparison. Physical fit, fatigue, friction and shipping performance remain untested unless separately evidenced. All model imagery uses the render skill and interactive geometry uses the P1S riser viewer. Outline thickness and depth bias are reduced consistently for these small, thin-walled parts to prevent hidden edges showing through solid surfaces; source geometry is unchanged.
Original prompt · private paths generalizeds · First Spark follow-up · Second Spark follow-up · Qwen tool-call deadline · Run manifest and hashes · Runtime versions · Dated Spark delivery status · Print-layout completeness audit