tinymixtral-v1.1-1b — 1B MoE

Earlier flagship (before v3.0). A wider sparse MoE: ~1.18B total / ~352M active params (top-2 of 8 routed experts), trained on a SmolLM-inspired blend.

HF: mikecovlee/tinymixtral-v1.1-1b

Model details

Parameter	Value
hidden_size	1024
num_layers	16
Attention	GQA (16 heads / 4 KV heads)
Head dim	64
RoPE theta	1,000,000
Norm	RMSNorm
Experts	8 routed (top-2)
Expert FFN	SwiGLU, intermediate = 2816
Vocab size	32,000
Max position	2,048
Total params	~1.18B
Active params	~352M

Training

Phase	Data	LR	Tokens	Steps	Time	End loss
Pretrain	SmolLM blend (89:11)	7e-4	4B	244,141	~102.5 h	1.9*
Continuation	non-overlapping new 4B, same 89:11	4.2e-4	4B	121,921	60.4 h	1.80

* Decay-phase loss plateaued at ~1.9 then spiked to 3.4 in the final ~300 steps as the data stream wrapped back to the start of the corpus.

Data: FineWeb-Edu + Cosmopedia v2 (89:11). Pretrain batch 16 × 1024 (16,384 tokens/step) → 4B tokens in 244,141 steps; warmup 2,000 steps. The continuation added 4B tokens in 121,921 steps (~32k tokens/step).

Evaluation

Measured with lm-evaluation-harness v0.4.12, 0-shot (cuda, bf16).

Task	Metric	1B MoE 4B	1B MoE 8B
HellaSwag	acc_norm	0.311	0.329
PIQA	acc	0.620	0.630
WinoGrande	acc	0.510	0.523
ARC-Easy	acc	0.463	0.479
ARC-Challenge	acc_norm	0.273	0.279
OpenBookQA	acc_norm	0.288	0.306
BoolQ	acc	0.548	0.620
LAMBADA	acc	0.200	0.234
Mean	—	0.402	0.425

Doubling pretrain tokens improved every task (BoolQ +7.2pp, LAMBADA +3.5pp, mean +2.3pp).

Few-shot: HellaSwag 10-shot acc_norm 0.312 · WinoGrande 5-shot acc 0.524 · ARC-Easy 25-shot acc_norm 0.468 · ARC-Challenge 25-shot acc_norm 0.262.

Comparison with similar models

Same suite and settings, measured locally (lm-evaluation-harness v0.4.12, 0-shot, cuda, bf16):

Task	Metric	1B MoE (8B)	SmolLM2-360M	Qwen3-0.6B
HellaSwag	acc_norm	0.329	0.563	0.473
PIQA	acc	0.630	0.719	0.673
WinoGrande	acc	0.523	0.587	0.563
ARC-Easy	acc	0.479	0.705	0.609
ARC-Challenge	acc_norm	0.279	0.383	0.340
OpenBookQA	acc_norm	0.306	0.372	0.316
BoolQ	acc	0.620	0.620	0.643
LAMBADA	acc	0.234	0.532	0.401

SmolLM2-360M was trained on 4T tokens and Qwen3-0.6B on 36T tokens, versus 8B tokens for the 1B MoE on a single consumer GPU; the gap is primarily a data-budget difference.

Instruction following (IFEval)

IFEval (instruction-level loose accuracy), --apply_chat_template:

Model	IFEval inst-level loose
1B MoE 4B	0.222
1B MoE 5B (post-train)	0.234
1B MoE 8B (continuation)	0.221

Instruction-following stays flat (~0.22) — more pretraining tokens do not improve it.

Post-training (negative result)

Post-trained from the lowest-loss pretrain checkpoint with the v1.1 recipe (Wiki + Cosmopedia v2 50:50, 1B tokens, lr 2e-5, warmup 300, WSD, 60,975 steps, ~26 h). Loss fell 2.98 → 1.8 but downstream metrics were unchanged:

Task	Pretrain (4B)	Post-train (5B)
HellaSwag (acc_norm)	0.311	0.313
PIQA (acc)	0.620	0.623
WinoGrande (acc)	0.510	0.505
ARC-Easy (acc)	0.463	0.465
ARC-Challenge (acc_norm)	0.273	0.272
OpenBookQA (acc_norm)	0.288	0.290
BoolQ (acc)	0.548	0.528
LAMBADA (acc)	0.200	0.195

At this scale, low-LR post-training on a knowledge blend lowers the loss but does not transfer to task ability.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("mikecovlee/tinymixtral-v1.1-1b", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("mikecovlee/tinymixtral-v1.1-1b")

Requires transformers and trust_remote_code=True (custom tinymixtral architecture).

Limitations

Base model only: no instruction tuning or safety alignment.
Trained on 8B tokens; English-centric. Superseded by the v3.0 flagship.

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 — this model (1B MoE, earlier flagship)
mikecovlee/tinymixtral-v1.1-0.5b — 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 the tinymistral spelling. Both belong to the same project.

Citation

@misc{tinymixtralv1.11b2026,
  title  = {TinyMixtral: a small Mixture-of-Experts language-model family},
  author = {Michael Lee},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/mikecovlee/tinymixtral-v1.1-1b}}
}

License

加载中…