tinymixtral-v2.0-beta — shared-expert experiment
Experimental architecture-ablation release: adds a DeepSeek-style always-on shared expert to the MoE layer. Beta quality; kept for reproducibility.
Model details
Architecture changes (vs v1.1)
| Parameter | Baseline (v1.1) | Shared Expert (v2) |
|---|---|---|
| Attention | GQA 14Q / 2KV (7:1) | GQA 16Q / 4KV (4:1) |
| Head dim | 64 | 56 |
| Experts | 6 routed (top-2) | 1 shared + 6 routed (top-2) |
| LR schedule | Cosine | WSD (Warmup-Stable-Decay) |
| Batch size | 24 | 22 |
| Total params | ~432M | ~498M |
| Active params | ~176M | ~241M |
The shared expert is always active (no routing), providing general-purpose features; routed experts specialize via top-2 gating. Output is the sum of both.
Training
- Data: FineWeb-Edu + Cosmopedia v2 blend (SmolLM-style).
- LR 7e-4, WSD schedule, warmup 2,000 steps, batch 22 × 1024, 4B tokens.
- Post-training: knowledge blend (Wiki + Cosmopedia v2), 1B tokens, lr 2e-5, warmup 300, WSD.
Evaluation
Measured with lm-evaluation-harness v0.4.12, 0-shot (cuda, bf16).
| Task | Metric | v1.1 (432M) | v2.0 (498M) |
|---|---|---|---|
| HellaSwag | acc_norm | 0.308 | 0.326 |
| PIQA | acc | 0.616 | 0.631 |
| WinoGrande | acc | 0.524 | 0.506 |
| ARC-Easy | acc | 0.456 | 0.474 |
| ARC-Challenge | acc_norm | 0.247 | 0.272 |
| OpenBookQA | acc_norm | 0.288 | 0.290 |
| BoolQ | acc | 0.606 | 0.455 |
| LAMBADA | acc | 0.227 | 0.224 |
| Mean | — | 0.409 | 0.397 |
Comparison with similar models
Same suite and settings, measured locally (lm-evaluation-harness v0.4.12, 0-shot, cuda, bf16):
| Task | Metric | v2.0 (498M) | SmolLM2-360M | Qwen3-0.6B |
|---|---|---|---|---|
| HellaSwag | acc_norm | 0.326 | 0.563 | 0.473 |
| PIQA | acc | 0.631 | 0.719 | 0.673 |
| WinoGrande | acc | 0.506 | 0.587 | 0.563 |
| ARC-Easy | acc | 0.474 | 0.705 | 0.609 |
| ARC-Challenge | acc_norm | 0.272 | 0.383 | 0.340 |
| OpenBookQA | acc_norm | 0.290 | 0.372 | 0.316 |
| BoolQ | acc | 0.455 | 0.620 | 0.643 |
| LAMBADA | acc | 0.224 | 0.532 | 0.401 |
SmolLM2-360M was trained on 4T tokens and Qwen3-0.6B on 36T tokens, versus 4B tokens for v2.0 beta on a single consumer GPU; the gap is primarily a data-budget difference.
Conclusion
The shared expert does not provide substantial improvement at this scale: mixed harness results (HellaSwag/ARC +1.5–2.5pp, BoolQ −15pp) for 15% more total params. At ~241M active params the routing overhead and representational fragmentation outweigh the capacity benefit. The design is more likely to pay off at 1B+ active params (cf. DeepSeek, Mixtral).
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("mikecovlee/tinymixtral-v2.0-beta", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("mikecovlee/tinymixtral-v2.0-beta")
Requires transformers and trust_remote_code=True (custom tinymixtral architecture).
Limitations
- Beta / experimental — released for the shared-expert ablation; not a maintained release.
- Base model only: no instruction tuning or safety alignment.
- Trained on 4B tokens; English-centric.
Family
mikecovlee/tinymixtral— 477.5M MoE base (v3.0 flagship)mikecovlee/tinymixtral-it— MoE instruction-tuned (v3.0-it, 3M SFT)mikecovlee/tinymistral-276m— 276M dense base (iso-active ablation)mikecovlee/tinymistral-276m-it— 276M dense instruction-tuned (3M SFT)mikecovlee/tinymixtral-v1.1-1b— 1B MoE (earlier flagship)mikecovlee/tinymixtral-v1.1-0.5b— 0.5B-class MoE (data-quality ablation)mikecovlee/tinymixtral-v2.0-beta— this model (shared-expert experiment, beta)mikecovlee/tinymixtral-v1.0— legacy (C4)
Naming. The MoE family is published under
tinymixtral; the dense 276M iso-active ablation companions use thetinymistralspelling. Both belong to the same project.
Citation
@misc{tinymixtralv2.0beta2026,
title = {TinyMixtral: a small Mixture-of-Experts language-model family},
author = {Michael Lee},
year = {2026},
howpublished = {\url{https://huggingface.co/mikecovlee/tinymixtral-v2.0-beta}}
}
License
MIT (Copyright (C) 2026 Michael Lee).