Does Parameter Count Predict
Construction-Domain Performance?

Diminishing returns at the frontier

Adding seven state-of-the-art giants — DeepSeek V4 Pro (1.6T), Xiaomi MiMo v2.5 Pro (1.02T), Kimi K2.6 (1T), GLM-5.1 (744B), MiniMax M2.7 (230B), DeepSeek V4 Flash (284B), Xiaomi MiMo (310B), plus the cloud high-ends of each family — does not break past the ceiling set by mid-size models. The standalone SOTA cluster lands between 74 and 90 average, the same band already occupied by Qwen 3.6 27B (91.3) and the cloud Gemini 3.5 Flash (91.1). The single clearest illustration: DeepSeek V4 Pro at 1.6 trillion parameters scores 85.7 — below its own 284B Flash sibling at 89.2. Kimi K2.6 (1T) scores 85.0; Xiaomi MiMo v2.5 Pro (1.02T) scores 83.5, below its own 310B base variant (85.4). The free GLM-5.1 (744B) reaches 90.1. The scaling charts make this visually explicit: the seven >300B asymptote lines all converge into the same narrow 85-91 band that the best sub-30B models already reach. For this construction benchmark, returns to scale flatten hard above roughly 25-40B active parameters, and total-parameter count beyond ~300B adds essentially nothing.

Parameter count is a weak predictor of performance

The largest model in the study (Qwen 3.6 35B-A3B, a 35 billion parameter mixture-of-experts variant with 3 billion active) averages 81.8 out of 100. The 27 billion dense sibling from the same family reaches 91.3. The 14 billion Ministral 3 dense model averages 72.5, while its 8 billion sibling scores 76.3. Within each family the curve is non-monotonic. Total parameter count alone explains less than half of the cross-model variance. Architecture, training mix, and reasoning support dominate raw parameter scaling.

Dense beats MoE at fixed active-parameter budget

Qwen 3.6 27B dense (91.3) outperforms the 35B-A3B mixture-of-experts variant (81.8) despite the latter having 30 percent more total parameters. Gemma 4 follows the same pattern: the 31B dense (83.0) and 26B-A4B (84.9) sit close to each other, but both trail Qwen 3.6 27B. For reasoning-heavy construction tasks the binding constraint appears to be active parameters, not total parameters. Mixture-of-experts variants pay a measurable accuracy cost on multi-step delay attribution and CPM reasoning.

Schedule generation universally overshoots

Task T3 asks the model to generate an 18-activity baseline schedule for a Cologne residential building from historical benchmark data. The golden answer arrives at 370 working days through a critical-path topology that uses start-to-start and finish-to-finish parallelism. Out of 16 models tested, only Gemma 4 E2B undershoots the golden duration (285 wd) and only Gemma 4 E4B comes within 35 percent (494 wd). The remaining models overshoot by factors ranging from 1.5 to 2.5, with most defaulting to long finish-to-start chains. This appears to be a structural weakness shared across all current base small language models: the models can list standard residential activities and can choose plausible durations from historical ranges, but they do not yet construct the parallel sub-schedules that real planners use.

CPM arithmetic is solvable; predecessor selection is not

Where T3 asks for schedule generation, CB3 provides a fixed 18-activity network and asks for the deterministic CPM forward and backward pass. The golden duration is 265 wd. Out of 16 models, six hit 265 exactly Qwen 3.6 27B, Qwen 3.6 35B-A3B, Qwen 3.5 9B, Qwen 3.5 4B, Gemma 4 26B-A4B, and Gemma 4 31B. These same models also correctly handle the SS+10 lag that makes Activity 8 critical with zero float, the rubric key discriminator for this task. The implication is that CPM computation on a given network is a tractable mechanical task for mid-size models, while predecessor-selection in schedule synthesis (T3) remains qualitatively harder.

Generational uplift within Qwen

Qwen 3.5 Flash, a hybrid linear plus sparse mixture-of-experts model with roughly 9 billion active parameters, averages 80.4. The next-generation Qwen 3.6 27B dense averages 91.3, an uplift of 10.9 points at similar effective compute. The mixture-of-experts variant from Qwen 3.6 (35B-A3B) gains only 1.4 points over the previous-generation flash. This suggests that the dense path in Qwen 3.6 captures most of the generational improvement, while the MoE path is closer to its predecessor in practical capability on this benchmark.

Reasoning capability has a lower size threshold

The Qwen 3.5 family spans 0.8B, 2B, 4B, 9B, and the cloud Flash variant. At 4B the model reasons cleanly through all six tasks and closes its think tags reliably. At 2B, reasoning still works but spirals on three of six tasks and recovers on the second attempt. At 0.8B, the thinking trigger spirals on all three attempts on the first task. After falling back to non-thinking mode the 0.8B model averages 36.2 out of 100, well below the 4B's 76.4. The practical reasoning floor for these construction tasks sits somewhere between 2B and 4B for Qwen-style models, and is lower for Matformer variants. Gemma 4 E2B at roughly 2 billion effective parameters completes all six tasks with thinking enabled and averages 74.75.

Curveball generalisation is uneven

CB1 (Finnish YSE 1998 contract, 21 clauses) is the hardest curveball. Only three models hit the rejection key discriminator on clause C-002 (lack of enforcement mechanism): Qwen 3.6 27B, Qwen 3.6 35B-A3B, and Nemotron 3 Nano 30B-A3B. Several mid-size models default to "Accepted" or "Requires Review" on clauses requiring modification under Finnish standards, suggesting that contract reasoning is heavily anchored in training data familiarity. CB2 (FIDIC offshore wind concurrent delay) is the easiest curveball: all sixteen models correctly classify the vessel breakdown as contractor-risk. The English-law Adyard concurrency principle is recognised by about half the field.

The champion model and the cost-effective alternative

Qwen 3.6 27B dense is the construction-domain champion at 91.3 of 100, hitting 9 of 10 key rubric discriminators across the six tasks. It achieves near-perfect contracts performance (96.5), strong delays handling (86.5), and reaches the ceiling on CPM analysis (98). Its only consistent weakness is the universal T3 schedule-generation overshoot. For local deployment on a 16GB consumer GPU, the strongest small model is Qwen 3.5 9B at 77.9 average; on Matformer architecture, Gemma 4 E4B at roughly 4 billion effective parameters reaches 77.6 with the fastest wall time of the local cohort. The Fine-Tuning study's headline that Qwen3-4B base hits 91 on construction tasks remains broadly consistent with the Qwen line's strong base performance demonstrated here.