A Guth Labs publication ↗
Back to News

TNG tests plug-in coding expertise without retraining a model's existing weights

TNG tests plug-in coding expertise without retraining a model's existing weights

November 10, 2026 | guthlabs.newsroom

TNG Technology Consulting's November 10 research account explores a way to specialize an open mixture-of-experts model while leaving its existing weights frozen. Its experiment adds independently trained Rust and Haskell experts to Qwen3.5-35B-A3B-Base.

TNG reports that combining the experts produces code-generation pass@1 scores of 0.70 for Rust and 0.50 for Haskell, against 0.62 and 0.32 for the base. These are developer-reported results, not independently replicated findings. The combined model retains most, not necessarily all, of each separate expert's gains. Additional router training improves the combination.

The publication says the resulting checkpoint needs no custom inference code. That does not mean zero deployment cost: it flags possible expert-parallelism and load-balancing penalties. Experiments cover two languages during continued pretraining; survival through a complete post-training pipeline and scaling to many experts remain open questions.

Why it matters for the fleet: analysis

This is a useful research direction for organizations that need several narrow capabilities without maintaining a completely separate model for every specialty. The practical question is not simply whether two benchmark scores improve. It is whether a combined model stays reliable on the organization's actual tasks, fits its serving budget, and can be replaced or rolled back predictably.

Before considering deployment, Guth should require a reproducible artifact, evaluate each specialty both alone and in combination, and measure latency and memory under its own workload. General tests should include unrelated tasks so a specialist gain does not conceal regression elsewhere. This is an evaluation proposal, not a report that Guth has adopted or tested the method.

Verification and evidence limits

The primary article was read directly, including its results and future-work sections. Its visible publication date is November 10; no exact publication time was available. No public release of these particular specialist checkpoints was established in this check. No independent replication was located.

Cited primary evidence

Independent corroboration

None established. Secondary summaries are not benchmark replication.

Additional background

None needed for this article.

Sources and citations

Revision history

Published .