pratikgx/launch-bench-results

Launch Bench results

These are my own benchmark runs of newly released open-weight models on one Apple Silicon Mac. Every run uses the same fixed harness and settings, so models can be compared over time. results.jsonl has one line per model run and will get a new line for each future model.

Hardware and software

• Apple M4 Pro (16 GPU cores), 24 GB unified memory, macOS 27.0.1
• mlx 0.32.4, mlx-lm 0.32.2, Python 3.12.14

Fixed settings (see harness/settings.json)

Speed

• Greedy decoding (temperature 0)
• Prompt lengths of 512, 2,048, 8,192 and 16,000 tokens. Each prompt is BOS plus the first N tokens of Pride and Prejudice (Project Gutenberg).
• 256 generated tokens with ignore_eos, prefill step 2048
• 1 warmup run, then 3 timed runs. Tables report the median.

Quality (run on the full-precision model, using seed-0 random subsets)

• GSM8K test, n=100, 0-shot chat. The answer is the last number in the reply, and replies cut off at 512 tokens count as wrong (strict). Run once with thinking off and once with thinking on.
• IFEval, n=100, thinking off, max 1,024 tokens, scored with the lm-eval IFEval instruction checkers (harness/ifeval_lib, Apache-2.0, Google Research)
• ARC-Easy test, n=200, lm-eval style 0-shot loglikelihood, no chat template
• HellaSwag validation, n=200, lm-eval style 0-shot loglikelihood, no chat template
• 3 sanity prompts

The subsets are small (100–200 items), so error bars are wide. For example, ±5–10 points at n=100. Treat small differences as noise.

Runs

2026-10-27: Laghu 152M (udgamlabs/Laghu-MLX)

Quality was measured on bf16. Speed was measured on bf16 and on a local 4-bit conversion (pratikgx/Laghu-MLX-4bit, q4, group size 64, 4.506 bits per weight).

Task	Result	Lab claim
GSM8K, thinking off	5.0%	–
GSM8K, thinking on	10.0%	7.9% (with thinking)
IFEval prompt-level strict / loose	0.29 / 0.30	26.1
IFEval instruction-level strict / loose	0.386 / 0.405	–
ARC-Easy acc / acc_norm	0.585 / 0.49	–
HellaSwag acc / acc_norm	0.305 / 0.38	–

On IFEval, 45% of replies hit the 1,024-token cap because the model tends to loop under greedy decoding. The lab's 49.9 average is for the base model across 5 tasks with a different setup, so it is not directly comparable.

Speed (median of 3):

Prompt tokens	bf16 decode	bf16 prefill	bf16 TTFT	bf16 peak mem	4-bit decode	4-bit prefill	4-bit peak mem
512	364 tok/s	11,516 tok/s	0.045 s	0.66 GB	402 tok/s	11,844 tok/s	0.38 GB
2,048	356 tok/s	11,264 tok/s	0.18 s	1.17 GB	409 tok/s	10,825 tok/s	1.11 GB
8,192	328 tok/s	9,198 tok/s	0.89 s	1.23 GB	397 tok/s	8,876 tok/s	1.15 GB
16,000	303 tok/s	7,554 tok/s	2.12 s	1.26 GB	379 tok/s	7,069 tok/s	1.15 GB

For comparison, the lab reports about 319 tok/s generation for bf16 on a Mac mini (Apple M6, 24 GB).

Files

• results.jsonl: one summary line per model run, with model, repo, hardware, settings and metrics
• runs/: raw per-run outputs. These are speed_*.json with every timed run, quality_*.json with per-item model replies and scores, and the run log.
• harness/: the scripts used. These are run_all.sh, bench_speed.py, bench_quality.py, make_subsets.py, summarize.py, common.py and settings.json.

The model outputs in runs/ come from the benchmarked models and are covered by those models' licenses.