vovaRL/NanoColibri-Instruct

Colibri-Nano

A 2.7B-parameter Mixture-of-Experts that runs on a 4 GB mini-PC — trained from scratch by volunteers passing a GPU baton, for about the price of a game console.

Parameters	2.7B total · 0.34B active per token (~12%)
Training tokens	~5.4B (pretrain + chat SFT)
Container	~1.2 GB int4, streams experts from disk
Compute	1 GPU at a time · ~50–100 H100-hours for the whole run
Cost	~$180–260 for the whole run
License	Apache-2.0 — code, checkpoints, and a ledger of who trained what

Architecture. 24 layers (1 dense + 23 MoE) · 64 experts, top-2 + a 2048-wide always-resident shared expert · GQA 4× · per-head QK-norm · sigmoid+bias router (DeepSeek-V3-style aux-loss-free balancing) · SmolLM2 tokenizer, 49 152 vocab · 4 096 context.

---

How it was built

1 — Relay pretraining. Pretraining is sequential, so contributors take turns: claim a lease on a HuggingFace repo (a real compare-and-swap, not a convention), pull the checkpoint, train a leg, push, hand the baton on. 20 000 updates ≈ 5.2B tokens of FineWeb-Edu, then a 1 500-update chat SFT.

2 — Engine-native. The architecture was chosen for the serving engine (hy3.c): int4 experts stream from disk under a fixed RAM budget, while the dense backbone and the shared expert stay resident. No conversion shims.

3 — The bet paid off. At a matched token budget, the MoE beats dense references that spend more compute per token.

---

Evaluation 1 — beats dense at the same token budget

Zero-shot accuracy (%). Colibri-Nano (2.7B total / 0.34B active) against two token-matched dense models.

Task	Colibri-Nano 0.34B act / 5.4B tok	Pythia-410M @ 6.3B tok (dense 405M)	Cerebras-GPT-256M @ 5.1B tok (dense)	Random
LAMBADA	26.3	26.3	29.3	0
PIQA	62.7	59.8	61.3	50
WinoGrande †	49.2	50.9	51.1	50
ARC-easy	43.4	41.3	41.0	25
ARC-challenge	22.8	18.8	17.0	25
OpenBookQA	22.0	15.6	15.8	25
HellaSwag	31.1	27.0	27.4	25

Identical prompts and scoring for all three columns (in-repo lm-eval-compatible harness; Cerebras row from arXiv:2304.03208 Table 8, same protocol). Colibri-Nano wins 5 of 7 against both token-matched dense models — ties LAMBADA, WinoGrande within noise — with fewer active parameters per token.

† WinoGrande differences are within noise (±1.4 pts).

‡ The Pythia checkpoint (step3000) is early in its LR schedule, which handicaps it somewhat — the fully-annealed Cerebras-GPT-256M column is the conservative comparison, and Nano wins that one too.

Evaluation 2 — quality vs training tokens

Average zero-shot accuracy over the 7 tasks above. Random-guess average on this task set: 28.6.

Model	Tokens (B)	Avg accuracy	Kind
Cerebras-GPT-111M	2.2	31.5	dense
Cerebras-GPT-256M	5.1	34.7	dense
Colibri-Nano	5.4	36.8	MoE (ours)
Pythia-410M @ step3000 ‡	6.3	34.2	dense
Cerebras-GPT-590M	11.8	37.1	dense
OPT-125M	180	37.1	dense
OPT-350M	180	39.5	dense
Pythia-160M	300	37.4	dense
Pythia-410M	300	41.9	dense

At 5.4B tokens, Colibri-Nano matches OPT-125M (33× the tokens) and Cerebras-GPT-590M (~4× the training compute), and clearly beats everything at its own budget. Models further down the table just ate more data.

Evaluation 3 — the efficiency frontier (context, not a target)

Different metric family. These are acc_norm numbers under the SmolLM2-card protocol and are not comparable with the accuracies in Evaluation 1–2.

Model	Active (M)	Total (M)	Tokens (B)	HellaSwag	ARC avg	PIQA	WinoGrande	OpenBookQA
Colibri-Nano	341	2700	5.4	35.6	32.8	62.7	49.2	29.6
SmolLM2-135M	135	135	2000	42.1	43.9	68.4	51.3	34.6
SmolLM2-360M	360	360	4000	54.5	53.0	71.7	52.5	37.4
Qwen2.5-0.5B	494	494	18000	51.2	45.4	69.9	54.1	37.4

The trillions-of-tokens ceiling for this size class — context, not a target. The gap is data volume (370–3300× more tokens), not architecture.

---

Reproducing every number on this card

python -m scripts.bench_nano --indir checkpoints/nano_chat

Every number here regenerates from the repo (docs/BENCH.md, docs/RESULTS.md).

Train it yourself: the one-line setup.sh + ./train.sh relay (see README in the training repo).

Limitations

• Undertrained by design of budget, not ambition: at ~5.4B tokens the model sits 370–3300× below the token counts of the frontier models in its size class, and Evaluation 3 shows the resulting gap.
• Speed benchmarks across target devices (tokens/s vs engine memory budget) are not yet published here; disk-pressure benchmarks (memory-capped cgroups, larger containers) are the successor project's first milestone.
• Small-model failure modes apply: factual errors, repetition, and weak multi-step reasoning.

What's next

A 7B chat+code pilot with RL post-training, then a 24–28B flagship whose container is deliberately bigger than a 16 GB laptop's memory — streaming experts from NVMe at 20+ tok/s (plan: docs/NEXT_MODEL.md).

---

License Apache-2.0 · HF repo vovaRL/NanoColibri-Instruct · engine: hy3.c