tinymixtral-v1.1-0.5b — 0.5B-class MoE
Dense-ish small MoE (~432M total / ~176M active, top-2 of 6 routed experts) trained on the SmolLM-inspired blend — the "data quality" ablation that replaced C4 with FineWeb-Edu + Cosmopedia v2.
The repository name says "0.5b" (storage/coarse size class); the trained model is ~432M total parameters, of which ~176M are active per token.
Model details
| Parameter | Value |
|---|---|
| Type | Sparse MoE, 6 routed experts (top-2) |
| Total params | ~432M |
| Active params | ~176M |
| Hidden size | 896 |
| Layers | 10 |
| Attention | GQA 14 Q / 2 KV heads (head dim 64) |
| Vocab | 32,000 |
| Context | 2,048 |
Training
- Data: FineWeb-Edu + Cosmopedia v2 ≈ 89:11 (sampling weights 36:4), 4B tokens.
- LR: 7e-4 (swept over {1e-4, 3e-4, 5e-4, 7e-4} on 100M-token runs; 7e-4 best).
- Batch: 24 × 1024; cosine schedule with 2,000-step warmup.
- Tokenizer: TinyLlama (32k SentencePiece).
Evaluation
Measured with lm-evaluation-harness v0.4.12, 0-shot (cuda, bf16).
| Task | Metric | v1.1 (432M) |
|---|---|---|
| HellaSwag | acc_norm | 0.308 |
| PIQA | acc | 0.616 |
| WinoGrande | acc | 0.524 |
| ARC-Easy | acc | 0.456 |
| ARC-Challenge | acc_norm | 0.247 |
| OpenBookQA | acc_norm | 0.288 |
| BoolQ | acc | 0.606 |
| LAMBADA | acc | 0.227 |
| Mean | — | 0.409 |
Comparison with similar models
Same suite and settings, measured locally (lm-evaluation-harness v0.4.12, 0-shot, cuda, bf16):
| Task | Metric | v1.1 (432M) | SmolLM2-360M | Qwen3-0.6B |
|---|---|---|---|---|
| HellaSwag | acc_norm | 0.308 | 0.563 | 0.473 |
| PIQA | acc | 0.616 | 0.719 | 0.673 |
| WinoGrande | acc | 0.524 | 0.587 | 0.563 |
| ARC-Easy | acc | 0.456 | 0.705 | 0.609 |
| ARC-Challenge | acc_norm | 0.247 | 0.383 | 0.340 |
| OpenBookQA | acc_norm | 0.288 | 0.372 | 0.316 |
| BoolQ | acc | 0.606 | 0.620 | 0.643 |
| LAMBADA | acc | 0.227 | 0.532 | 0.401 |
SmolLM2-360M was trained on 4T tokens and Qwen3-0.6B on 36T tokens, versus 4B tokens for v1.1 (~1000× less) on a single consumer GPU — the gap is primarily a data-budget difference.
The C4 → SmolLM data-quality switch mainly improved ARC-Easy (+3.4pp) and BoolQ (+2.7pp) versus the legacy v1.0 C4 model (mean 0.403 → 0.409); ARC-Challenge was unchanged (0.247).
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("mikecovlee/tinymixtral-v1.1-0.5b", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("mikecovlee/tinymixtral-v1.1-0.5b")
Requires transformers and trust_remote_code=True (custom tinymixtral architecture).
Limitations
- Base model only: no instruction tuning or safety alignment.
- Trained on only 4B tokens; English-centric. Superseded by v3.0.
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— this model (0.5B-class MoE, data-quality ablation)mikecovlee/tinymixtral-v2.0-beta— 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{tinymixtralv1.10.5b2026,
title = {TinyMixtral: a small Mixture-of-Experts language-model family},
author = {Michael Lee},
year = {2026},
howpublished = {\url{https://huggingface.co/mikecovlee/tinymixtral-v1.1-0.5b}}
}
License
MIT (Copyright (C) 2026 Michael Lee).