LittleKedi-medium-wide-sparse-base

LittleKedi-medium-wide-sparse-base is a small DeepSeek-V3-style Mixture-of-Experts language model trained from scratch. It has 305.4M parameters in total, of which 31.7M are active per token (23.29M excluding the embedding table). It was pretrained on a 20,400M-token mix (FineWeb-Edu, Cosmopedia v2 (SmolLM-Corpus), FineMath 4+, Reddit comments, github-code-clean, SmolTalk, Wikipedia (English, 2023-11-01)).

Part of the LittleKedi family. Not affiliated with DeepSeek.

This is the base model, before any instruction tuning. It continues text; it doesn't follow instructions. It's an educational and research artifact: it writes fluent, on-topic English but knows few facts and confidently makes things up. Don't use it for anything that matters.

The LittleKedi family

LittleKedi is a family of small DeepSeek-V3-style Mixture-of-Experts language models trained from scratch. Every size comes in three variants that differ in how sparse they are: how many routed experts they have, and what share of their parameters a token actually passes through.

Variant	Routed experts
standard (<size>)	a few wide experts; the densest variant
sparse (<size>-sparse)	many narrower, fine-grained experts, more of them active per token
wide-sparse (<size>-wide-sparse)	more experts again; the largest total and smallest active share

More experts means more total capacity to store knowledge, but each expert sees fewer tokens (tokens × top-k ÷ experts per layer), so sparser variants need more training data before their experts are well trained. The variants of a size share a tokenizer and dataset, so they can be compared directly. Shapes change as the family scales; the table below is generated from the current presets.

This card is for medium-wide-sparse, the wide-sparse medium model: 31.7M of its 305.4M parameters (10%) are active per token. Its siblings are medium and medium-sparse.

Model	Variant	Vocab	Total	Active	Active non-emb.	Active %	Routed experts
tiny	standard	8K	2.0M	1.5M	0.5M	78%	8 × 64, top-2
tiny-sparse	sparse	8K	2.6M	1.6M	0.5M	60%	32 × 32, top-4
tiny-wide-sparse	wide-sparse	8K	3.8M	1.6M	0.5M	42%	64 × 32, top-4
mini	standard	8K	9.4M	4.8M	3.2M	51%	12 × 128, top-3
mini-sparse	sparse	8K	19.8M	4.8M	3.2M	24%	64 × 64, top-6
mini-wide-sparse	wide-sparse	8K	36.4M	4.9M	3.3M	13%	128 × 64, top-6
small	standard	8K	12.2M	6.3M	4.2M	52%	16 × 128, top-4
small-sparse	sparse	8K	24.1M	6.4M	4.3M	27%	80 × 64, top-8
small-wide-sparse	wide-sparse	8K	43.9M	6.5M	4.4M	15%	160 × 64, top-8
moderate	standard	16K	45.5M	18.9M	12.6M	42%	24 × 192, top-4
moderate-sparse	sparse	16K	93.5M	19.1M	12.8M	20%	120 × 96, top-8
moderate-wide-sparse	wide-sparse	16K	173.4M	19.4M	13.1M	11%	240 × 96, top-8
medium	standard	16K	78.2M	31.0M	22.6M	40%	24 × 256, top-4
medium-sparse	sparse	16K	163.4M	31.3M	22.9M	19%	120 × 128, top-8
medium-wide-sparse ← this model	wide-sparse	16K	305.4M	31.7M	23.3M	10%	240 × 128, top-8
base	standard	32K	258.7M	105.4M	80.2M	41%	32 × 256, top-6
base-sparse	sparse	32K	542.8M	106.3M	81.2M	20%	160 × 128, top-12
base-wide-sparse	wide-sparse	32K	1015.9M	107.6M	82.4M	11%	320 × 128, top-12

"Active" counts everything a token passes through: embeddings, attention, the dense layer, shared experts and its top-k routed experts. "Active non-embedding" leaves out the embedding table (tied with the output head), which costs memory but almost no compute; it's the fair number for comparing compute between models.

Architecture

Component	This model
Parameters	305.36M total, 31.68M active per token (10%), 23.29M active non-embedding
Layers / hidden size	8 / 512 (the first 2 layers have a dense SwiGLU FFN of 1408; layers 2–7 are MoE)
Attention	Multi-head Latent Attention (MLA): 8 heads, KV latent 128, query latent 256, decoupled RoPE dim 32, head dims 64+32 (q/k) and 64 (v). The KV cache stores only the 128-dim latent plus a 32-dim RoPE key per token
MoE	DeepSeekMoE: 240 routed experts (SwiGLU, 128 hidden), top-8, plus 2 shared experts; sigmoid gating; group-limited routing (8 groups, top-4)
Load balancing	Auxiliary-loss-free bias balancing (update speed 0.001) plus a small sequence-wise loss (α = 0.0001). Over the second half of training the median batch put 1.92× the mean load on the busiest expert (2.28× in the worst layer), and 0.26% of expert assignments were dropped for capacity.
Expert dispatch	Training: fixed capacity (capacity factor 2.0; overflow assignments are dropped). Inference is always dropless
Embeddings	Input embedding and output head tied
Context	2,048 tokens

Training

Data	20,400M tokens, 27,145,344 documents (~67 tokens per parameter, ~644 per active parameter), one pass, sources interleaved throughout (mix below)
Tokenizer	Byte-level BPE, 16,384 tokens (~4.00 bytes/token on this mix); digits split individually; chat special tokens built in
Steps	311,279 × 65,536 tokens (batch 16 × 2048 × 2 accumulation)
Optimizer	Muon (momentum 0.95, Nesterov, 5 Newton-Schulz steps; each expert's matrix orthogonalized separately) for hidden matrices, AdamW (β = 0.9, 0.95) for embeddings, norms and the router; weight decay 0.1, gradient clipping 1.0
Learning rate	2e-3 peak; WSD: 200 warm-up steps, constant, then linear decay to 0 over the last 20% of steps (schedule set for 311,279 steps)
Precision	bf16 autocast, torch.compile
Compute	~2.9e18 FLOPs (6 × active non-embedding parameters × tokens)
Hardware	One rented NVIDIA GPU at a time on RunPod (Secure cloud), resumed across several pods; the second half on an A40 and RTX 3090s: about 76.7 hours at a median ~73.9K tokens/s

Source	What it is	Tokens	Share
FineWeb-Edu sample/100BT	educational web pages	11,820M	58%
Cosmopedia v2 (SmolLM-Corpus)	synthetic textbooks and stories	2,446M	12%
FineMath 4+	maths web pages	1,875M	9%
Reddit comments	forum comments	1,630M	8%
github-code-clean	source code	958M	5%
SmolTalk	chat conversations	917M	4%
Wikipedia (English, 2023-11-01)	encyclopedia articles	754M	4%

Evaluation

Metric	Value
Validation loss (nats/token, held-out mix)	2.340
Perplexity	10.4
Bits per byte	0.854
Validation loss: code	1.160
Validation loss: cosmo-v2	1.823
Validation loss: finemath	1.842
Validation loss: fineweb-edu-100bt	2.641
Validation loss: reddit	2.907
Validation loss: smoltalk	1.344
Validation loss: wikipedia	2.604
Experts never chosen on the validation sample	0

Validation loss over training: 5.05 (16M tok) → 2.78 (2,933M tok) → 2.75 (5,849M tok) → 2.74 (8,766M tok) → 2.73 (11,665M tok) → 2.73 (14,582M tok) → 2.66 (17,498M tok) → 2.34 (20,400M tok)

Zero-shot benchmarks (evals.py)

Task	acc_norm	acc	Items	Chance
SciQ	72.2% ± 1.4%	78.8%	1,000	25%
ARC-Easy	47.1% ± 1.0%	53.2%	2,376	25%
HellaSwag	36.0% ± 0.5%	31.3%	10,042	25%

Each answer choice is scored by the log-probability of its text after the prompt, using lm-evaluation-harness prompts (SciQ includes its supporting passage). acc_norm divides by the answer's length in characters; acc doesn't. Splits: SciQ test, ARC-Easy test, HellaSwag validation.

MMLU (mmlu_eval.py)

Format	Accuracy	Questions scored	Average few-shot examples
letter	24.5%	14,042 (skipped 0)	4.9
cloze	27.8%	14,042 (skipped 0)	0.0

Category	Subjects	Questions	Letter	Cloze
STEM	19	3,153	23.8%	25.9%
Humanities	13	4,705	25.1%	26.1%
Social Sciences	12	3,077	23.2%	29.9%
Other	13	3,107	25.5%	30.4%

Biology & medicine (cloze): 29.9% on 2,094 questions, +4.9 points against 25% chance (about 5.2 standard errors, significant). This pools the 10 life-science and health subjects; it isn't an official MMLU category.

In the letter format the model compares " A"–" D", with same-subject examples added while they fit the context. In the cloze format each answer's text is scored by log-probability per byte; at this size the cloze results are the meaningful ones. Chance is 25%.

Compared with its siblings and other small models

	This model	moderate-wide-sparse	Pythia-70M	GPT-2	Supra-50M
Parameters (total)	305.4M	173.4M	70.4M	124.4M	51.8M
Active non-embedding parameters	23.3M	13.1M	18.9M	85.1M	35.4M
Training tokens	20,400M	12,500M	300B (the Pile)	undisclosed (WebText)	20B (FineWeb-Edu)
SciQ (acc_norm)	72.2%	68.6%	55.2%	53.2%	77.2%
ARC-Easy (acc_norm)	47.1%	45.2%	35.0%	42.0%	52.2%
HellaSwag (acc_norm)	36.0%	32.8%	–	29.5%	31.8%
MMLU (cloze)	27.8%	27.4%	–	–	–
MMLU biology & medicine (cloze)	29.9%	29.1%	–	–	–
Validation loss (held-out mix)	2.340	2.478	–	–	–

LittleKedi columns are measured with this repo's evals.py and mmlu_eval.py on the checkpoints named. Validation losses are only comparable between models that share a tokenizer and validation set. The other columns are published figures: Pythia-70M from EleutherAI's evaluations as quoted on the Wisp-15M card, GPT-2 and Supra-50M from the Supra-50M-Base card. Their prompts may differ from ours, so treat gaps of a few points as noise.

What it can and can't do

The sparsest medium model: about 23.3M active non-embedding parameters per token, spread over 240 narrow experts (305M parameters in total), trained on 20.4B tokens. It's the strongest LittleKedi so far: validation loss 2.340 (moderate-wide-sparse: 2.478; the validation sets differ slightly), and better than moderate-wide-sparse on every benchmark. The gains are about what ~2.9x the training compute should buy, not a jump in kind: it gets the most common facts right reliably now, but it's still a model of the form of text more than its content, and specific facts, numbers and code are usually wrong. A research artifact for studying sparsity in small MoE models, not a source of information.

Most of its quality arrived in the last 20% of training. Validation loss sat on a plateau of ~2.73 for most of the run and fell by ~0.39 during the learning-rate decay, about three times the usual cooldown gain, which suggests the peak learning rate (2e-3) was higher than it needed to be.

What it does

• ✅ Writes fluent, well-organised paragraphs in the register of textbooks, encyclopedias and educational web text, staying on topic for a paragraph or more.
• ✅ Gets the most common facts right reliably: "The French Revolution began in" → "1789", "World War II ended in" → "1945" and "The first person to walk on the Moon was" → "Neil Armstrong" (each 3 of 3 seeds; the Moon landing's "1969" on 2 of 3), "The capital of France is" → "Paris" (2 of 3), "Photosynthesis is the process by which" → "plants convert light energy into chemical energy" (2 of 3), "The heart pumps" → "blood through the body to supply oxygen and nutrients to all cells". Einstein is "born in 1879" on 2 of 3 seeds and "a German-born physicist" on the third.
• ✅ Occasionally writes correct short code: one seed of "def is_prime(n):" produced the standard trial-division loop up to int(n**0.5) + 1. It knows the shape of common functions (docstrings, base cases, recursion) even when the logic is wrong (below).
• ✅ The best benchmark scores in the family (zero-shot): SciQ 72.2% (acc_norm; 78.8% acc), ARC-Easy 47.1% (53.2% acc), HellaSwag 36.0%, PIQA 64.7%, ARC-Challenge 26.3%, ArithMark-3 37.8%, up from 68.6 / 45.2 / 32.8 / 62.0 / 25.4 / 36.2 for moderate-wide-sparse. On the Open SLM Leaderboard's Intelligence Index that's ~19.4, against ~15.9 for moderate.
• ✅ Above chance on MMLU in the cloze format: 27.8% overall, ~30% on the social sciences and "other" categories. STEM (25.9%) and the humanities (26.1%) are still at about chance, and the multiple-choice letter format is at chance too (24.5%).

What it can't do

• ❌ Less common facts. "The chemical symbol for gold is" → "Cu" or "Pt"; "The largest planet in the solar system is" → "Ganymede", "the Earth" or a comparison with Mercury; "Water boils at" → "100°C" on one seed, "100 degrees Fahrenheit" and "103.6 degrees Fahrenheit" on the others. Around a right answer the details are made up: Einstein "born in Budapest, Hungary" or "in St. Petersburg, Russia", France "ruled by King Louis XVI… a monarchy under Napoleon III". Right and wrong answers come with the same confidence.
• ❌ Arithmetic beyond the very simplest. "2 + 2 =" → "4" on 2 of 3 seeds ("5" on the third), but "7 + 6 =" → "14", "15" or a multiple-choice question whose marked answer is "45".
• ❌ Working code, most of the time. Its Fibonacci is wrong on every seed ("return 2 * fibonacci(n - 1) + fibonacci(n - 2)", or the triangle-number formula), and two of three is_prime completions are just "return n % 2 == 0".
• ❌ Instructions or questions. It's a base model: it continues text. Asked "What is the largest planet in the solar system?" in a User: / Assistant: exchange, it repeats the question, writes a quiz page, or continues a dialogue that names Mercury or Pluto.
• ❌ Long-range coherence. A paragraph holds together; beyond that it drifts and repeats.

Decoding

Greedy decoding still loops, though later than the smaller models ("The heart is a muscle that pumps blood throughout the body. It is a muscle that is made up of two parts… The heart is a muscle that is made up"). Use sampling with temperature ≈ 0.7, top-p ≈ 0.9 and a repetition penalty of about 1.15.

Usage

PyTorch (needs the littlekedi package from the training code):

import json, torch
from safetensors.torch import load_file
from littlekedi import LittleKediModel, ModelConfig
from littlekedi.tokenizer import BPETokenizer

model = LittleKediModel(ModelConfig.from_dict(json.load(open("config.json")))).eval()
model.load_state_dict(load_file("model.safetensors"))
tok = BPETokenizer.from_file("tokenizer.json")
ids = torch.tensor([[tok.eot_id] + tok.encode("Photosynthesis is")])
print(tok.decode(model.generate(ids, 60, temperature=0.7, top_p=0.9, repetition_penalty=1.15)[0].tolist()))

Or interactively: python sample.py --ckpt <checkpoint>.pt. Base models are released as PyTorch weights only; GGUF builds come with the instruct models.

Files

File	Contents
model.safetensors	PyTorch weights (embedding tied with the output head), littlekedi parameter names
config.json	ModelConfig for littlekedi
tokenizer.json	Byte-level BPE tokenizer (Hugging Face tokenizers format)
little_kedi.jpg	The LittleKedi logo

Acknowledgements

• The architecture follows DeepSeek-V3 (DeepSeek-AI, 2024): MLA, DeepSeekMoE and auxiliary-loss-free balancing.
• Muon (Keller Jordan et al., 2024) with Moonlight's update scaling (Liu et al., 2025).
• Pretraining data: FineWeb-Edu (ODC-BY 1.0).
• Pretraining data: Cosmopedia v2 (ODC-BY 1.0).
• Pretraining data: FineMath 4+ (ODC-BY 1.0).
• Pretraining data: Reddit comments (no license stated on the dataset card; the comments are Reddit users' content).
• Pretraining data: github-code-clean (Apache 2.0, with each file under its own open-source license).
• Pretraining data: SmolTalk (Apache 2.0 for its new subsets, with the datasets it includes under their own licenses).
• Pretraining data: Wikipedia (CC BY-SA 3.0 and GFDL).

Card generated 2026-10-30 by prepare_release.py from medium-wide-sparse.pt (step 311278).