Qwen3.6-35B-A3B → A6B: Architecture and Tuning

Doubling the active experts on a frozen 35B checkpoint, then healing the resulting mixture with a residual-delta expert fine-tune. Same weights, wider router.

Base model Qwen3.6-35B-A3B · 40 transformer layers · 256 routed experts + 1 shared expert per MoE layer · measured 2026-07

1 · What changes

The only structural edit is the router's top-k. Everything downstream reacts to it.

Stock A3B — router top-k = 8 token hidden state Router — softmax over 256 experts (fp32) → keep top-8 → renormalize 256 routed experts · 8 activated per token Σ combine (renorm) shared expert block output → next layer ≈3B active / token Our A6B — router top-k = 32 token hidden state Router — softmax over 256 experts (fp32) → keep top-32 → renormalize 256 routed experts · 32 activated per token Σ combine (renorm) shared expert block output → next layer ≈6.6B active / token identical weights · only top-k changes
  • Top-8 routed experts — stock selection (46% of gate mass)
  • Added ranks 9–32 — untrained mixture members, 54% of gate mass
  • Inactive experts (weights present, not activated)
  • Shared expert (always active)

The router computes a softmax over all 256 experts in fp32, keeps the top-k, and renormalizes the selected weights unconditionally — renormalization is baked into the architecture, so going from k=8 to k=32 silently redistributes the entire gate mass across four times as many experts. Because the added ranks 9–32 carry 54.0% of that renormalized mass, they are load-bearing, not tail noise. Experts are stored as packed 3D tensors (gate_up_proj [256, 1024, 2048], down_proj [256, 2048, 512] per layer), so widening k adds compute but zero new parameters.

Measured risk — naive k=32 without retraining. Flipping top-k alone degrades knowledge benchmarks: MMLU 0.843 → 0.807 and GSM8K 0.893 → 0.865 (n=600). The 24 added experts per token were never trained to collaborate inside the mix — an uncalibrated mixture. That is exactly what the tuning recipe below repairs.

2 · Why expand k

Roughly 2× active compute per token, at zero weight cost.

Widening the router turns latent capacity that already lives in the checkpoint into active compute: ≈3B → ≈6.6B active parameters per token with no new weights to store or serve. The working hypothesis is a complexity gradient — complex, multi-step work (agentic terminal use, hard coding) should benefit most from the extra mixture capacity, while simple factual recall benefits least. So the payoff and the risk sit on different axes: reasoning throughput goes up, and knowledge is the axis we must actively protect.

Coding (post-train)

0.902

HumanEval vs stock-k8 0.866 · n=164, trend

Knowledge (to heal)

−2.0 pt

MMLU vs stock · statistical tie, p=0.14

Active compute

≈2×

≈6.6B active/token · +0 weights

Gains grow with task complexity; the knowledge gap is currently a statistical tie rather than a clean win, so MMLU is the axis being actively healed rather than a settled result.

3 · The tuning recipe

Profile the wider router, then teach only the experts it leans on to collaborate.

a

Router profiling

Run the target corpus through the model at k=32 and count token-level routing frequency for every expert.

output: per-expert routing histogram, all 40 layers
b

Expert selection

Keep experts by cumulative routing frequency up to top-p 0.2 — the experts that actually do the work.

833 of 10,240 layer-experts40 layers × 256
c

Residual-delta ESFT

Train only delta tensors added to the selected experts' FFN slices. The router and every other weight stay frozen. Deltas are toggleable — patched on or off at load time.

2.62B trainable of 35B (7.5%) delta on/off = same file 5.2 GB artifact
d

Data

Mixed corpus, 428M tokens, contamination-gated to zero residual hits against every eval benchmark. A small general + knowledge replay slice is deliberate forgetting insurance.

  • Agentic terminal trajectories 62%
  • Coding 12%
  • Tool-calling 11%
  • Math 10%
  • SWE-RL format envelope 2.2%
  • General + knowledge replay 2.8%
e

Training

8×GPU DDP at sequence length 7168, with fused linear cross-entropy (no logits materialization) and Adafactor.

8×GPU DDPseq 7168fused linear CEAdafactor≈19h

4 · Why the router stays frozen

The routing distribution is the coordinate system the deltas are trained in.

Fixed coordinate system

The deltas are learned relative to a specific routing distribution. Freeze it and every delta keeps meaning the same thing at inference; move it and the ground shifts underneath them.

No routing collapse

Letting the router move under SFT/RL is the classic path to routing collapse, where a few experts capture all traffic. A frozen router removes that failure mode entirely.

Exactly recoverable

Because nothing but the deltas changed, turning the delta off returns the exact stock model. The base checkpoint is never at risk.

5 · Roadmap

From supervised mixture-healing toward Terminal-Bench.

Done

Mixed SFT

Residual-delta ESFT on the 428M mixed corpus.

Then

GRPO RL

Rule-based patch-similarity reward.

Target

Terminal-Bench

The north-star agentic benchmark.