description: Models spill runs, with sizes and measured pass times: Qwen2.5 0.5B to 32B and Llama 3.3 70B in bf16 on a 48 GB Mac, plus supported architecture families.

Models and architectures

Curated tags

tag	params	family (state)	bf16	8-bit	on 48 GB	disk needed
qwen2.5:0.5b	0.5B	Qwen2/2.5 (verified)	0.9 GB	0.7 GB	resident	21 GB+
qwen2.5:1.5b	1.5B	Qwen2/2.5 (verified)	2.9 GB	1.5 GB	resident	23 GB+
qwen2.5:7b	7B	Qwen2/2.5 (verified)	14.2 GB	7.5 GB	resident	34 GB+
qwen2.5:32b	32B	Qwen2/2.5 (verified)	61 GB	33 GB	streamed	81 GB+
llama3.3:70b	70B	Llama 3.x (verified)	131 GB	70 GB	streamed	151 GB+

For a project made by spill init, the student defaults to qwen2.5:0.5b on every machine, the smallest curated model, so a first result takes minutes; spill plan states the reason for this machine and what else it could train. There is no default teacher and no default judge: you name one (--teacher, --judge), or accept the judge spill plan recommends for the machine (--judge recommended). Each of the three roles takes any curated tag above or any Hugging Face repo id with a supported architecture (spill run org/name file.jsonl); an unsupported architecture is refused in one line naming its family. The older folder workflow (spill build <folder>) still defaults to qwen2.5:7b and llama3.3:70b on Apple silicon. spill models prints the live table for your machine. Pinned revisions with every file's sha256 are in streamweights/models.yaml (qwen2.5:0.5b and qwen2.5:1.5b share one tokenizer, which is what logit distillation needs).

Disk: each model needs its size plus a 20 GB floor that downloads never cross. Downloads check the disk before the first byte, show one line with speed and ETA, and resume after an interruption.

Architecture families

spill models --architectures prints the live table.

state	families
verified	Llama 3.x, Qwen2/2.5, Qwen3 (dense), Phi 3/4, Gemma 2, Mistral
expected	Gemma 3 (text): per-layer alternating sliding-window attention needs per-layer cache windows; not wired yet
not_yet	Mixtral, DeepSeek V2/V3, Qwen MoE, Llama 4: mixture-of-experts, where per-token expert routing defeats layer-order streaming. Multimodal models (anything with a vision tower): vision towers are not streamed yet (planned, see plan.md). Mamba: state-space recurrence has no KV cache to batch around. Jamba: recurrent layers break the per-layer stream loop

"Verified" means the streamed output was tested identical to in-memory execution on 20 prompts, plus first-token agreement with an independent mlx-lm implementation. A family moves from expected to verified when a small model of that family passes both tests. Shared-prefix reuse is wired for the dense families that take a causal mask string (all the verified families except Gemma 2).

Measured on a 48 GB M4 Pro

Eval decode on a streamed model is disk-bound; prefill and training are compute-bound (2 and 6 times parameters times tokens, divided by achieved FLOP/s). Per-phase numbers live in the reports under docs/reports.

model	quant	placement	workload	pass time	measured
qwen2.5:0.5b	bf16	resident	2,000-row eval	n/a	712 tok/s aggregate (report 002)
llama3.3:70b	bf16	streamed	first streaming runner	31.6 s	(report 002)
llama3.3:70b	8-bit	streamed	golden path	18.8 s	(report 002)
llama3.3:70b	bf16	streamed	1,334 short and summarize rows	33-38 s	15 h 35 m unattended (report 007)
llama3.3:70b	bf16	streamed	666 one-thousand-token rows	33-37 s clean	23 h 51 m, zero interventions (report 007)
llama3.3:70b	bf16	streamed	LoRA training, 3 steps	417 s per step	4.55 TFLOP/s, 11 tokens/s (report 008)