pratikgx/1000-rewrites-later

1,000 Rewrites Later

What happens when a language model rewrites its own output, over and over? Eleven open-weight models were each given the Gettysburg Address and asked to rewrite their own previous output 1,000 times, all running locally on one Mac mini (Apple M5 Pro, 64 GB unified memory). This dataset contains every round's text plus speed and meaning-drift metrics, for 19 model-runs and 19,019 rows.

Author: Pratik Gandhi (@pratikg on X). Collected Oct 23–25, 2026.

Runs

run (split)	models	temperature	rounds	rows	notes
main	9	0.7	1,000 each	9,009	all models fully in unified memory
hot	8 (all of main except Qwen3.8 27B)	1.0	1,000 each	8,008	same as main, hotter sampling
frontier	2	0.7	1,000 each	2,002	models far larger than 64 GB; MoE experts streamed from the internal SSD

Each model-run has 1,001 rows: round 0 is the original text (status = "seed"), and rounds 1–1000 are the rewrites. No rounds are missing. All 19,000 rewrites ended with finish_reason = "stop" (none hit the token limit), with 0 retries and 0 flagged preambles.

Setup

• Hardware: Mac mini, Apple M5 Pro, 64 GB unified memory, macOS 27.0.1. Stock GPU wired-memory limit (not raised). Python 3.12.
• Seed text: the Gettysburg Address (Bliss copy, 272 words by the analysis tokenizer, 278 by the runner's counter). Public domain. Included as seed_gettysburg.txt.
• Prompt (every round, no system prompt, no history):

  Rewrite the following text in your own words. Keep the same meaning and roughly the same length. Output only the rewritten text, with no title, no preamble and no notes.
  
  Text:
  {text}
  

  {text} is the previous round's output. Round 1 uses the original.
• Sampling: temperature 0.7 (main, frontier) or 1.0 (hot), top_p 0.95, seed 42, max 900 new tokens, thinking/reasoning disabled.
• Runtimes:
  • mlx: mlx-lm (MLX 0.32.x) in-process. mlxvlm: mlx-vlm 0.7.10 (MLX 0.32.4) in-process. Weights fully resident in memory.
  • sushi: sushi 1.3.0 server (MLX 0.32.3) behind an OpenAI-compatible API, ctx 8192, SSD budget 20 GiB, experts streamed from the SSD.
  • omlx: oMLX server with MoE expert offload (48 of 384 routed experts resident per layer, the rest streamed from the SSD), OpenAI-compatible API.

Models

model	repo	quant	runtime	SSD-streamed	download (GB)	runs
Qwen3.5 0.8B	mlx-community/Qwen3.5-0.8B-4bit	4bit	mlx	no	0.7	main, hot
DeepSeek-V2-Lite	mlx-community/DeepSeek-V2-Lite-Chat-4bit-mlx	4bit	mlx	no	8.8	main, hot
Bonsai 2 9B	prism-ml/Ternary-Bonsai-2-9B-mlx-2bit	2bit (ternary)	mlx	no	3.5	main, hot
Qwen3.6 35B-A3B	mlx-community/Qwen3.6-35B-A3B-4bit	4bit	mlx	no	20.4	main, hot
Gemma 4 26B	mlx-community/gemma-4-26b-a4b-it-4bit	4bit	mlx	no	15.4	main, hot
GLM-4.7 Flash	mlx-community/GLM-4.7-Flash-4bit	4bit	mlx	no	16.9	main, hot
Qwen 3.8 Flash Next (pruned)	neopolita/Niwaki-Qwen3.8-Flash-Next-141B-A6B-v4.0-mlx	mixed (see repo)	mlxvlm	no	44.8	main, hot
Qwen3.8 27B	mlx-community/Qwen3.8-27B-4bit	4bit	mlx	no	16.1	main
Qwen3-Next 80B	mlx-community/Qwen3-Next-80B-A3B-Instruct-4bit	4bit	mlx	no	44.9	main, hot
GLM-5.3 Flash	beamster/GLM-5.3-Flash-Sushi-2bpw	2bpw	sushi	yes	87.6	frontier
DeepSeek V4.1 Flash	mlx-community/DeepSeek-V4.1-Flash-1017-oQ2.5e-mtp	oQ2.5e (~2.5-bit) + MTP	omlx	yes	342.7	frontier

Metrics (how they were computed)

These are the same numbers used in the project's charts and video (analyze.py is included).

• sim_to_original: cosine similarity between the round's text and the original, using sentence-transformers/all-MiniLM-L6-v2 normalized embeddings. Unrelated text can score slightly below 0.
• sim_to_prev: the same cosine similarity, measured against the previous round (the size of each step).
• same_meaning: sim_to_original >= 0.75, the "same-meaning line."
• words_analysis, unique_ratio, rep_trigram: word count, unique words ÷ words, and the share of repeated word trigrams, using the regex tokenizer [A-Za-z0-9']+ on lowercased text. words is the runner's own count, which is slightly different.
• Collapse round (in summary): the first round of the first 3-round streak where any of these is true: sim_to_original < 0.75, rep_trigram > 0.5, words < 30% of the seed, or a non-ok status. null means the model held the meaning for all 1,000 rounds.
• tok_s is generation tokens/s and prompt_tok_s is prefill tokens/s, both as reported by the runtime. seconds is wall time per round, including prefill. ttft_s is time to first token. peak_mem_gb is the MLX peak memory for in-process runs (null for the server-based frontier runs).
• ts: timestamp in Pacific time (ISO 8601 with offset).

Columns

Per-round splits (main, hot, frontier): run_type, model_label, run_label, model_name, hf_repo, quant, runtime, engine, ssd_streamed, temperature, top_p, round, text, words, words_analysis, sim_to_original, sim_to_prev, same_meaning, unique_ratio, rep_trigram, token_count, prompt_tokens, tok_s, prompt_tok_s, seconds, ttft_s, peak_mem_gb, finish_reason, retries, preamble, status, ts.

summary has one row per model-run with: collapse round, held_meaning, final/min similarity, rounds at or above the line, mean/median tok/s, median s/round, total hours, total generated tokens, peak memory, seed/final word counts, first/last timestamps, final text, prompt, and machine info.

Key findings

model	collapse round, temp 0.7	collapse round, temp 1.0	final similarity (0.7)	mean tok/s (0.7)
GLM-5.3 Flash (SSD)	never	–	0.912	5.8
DeepSeek V4.1 Flash (SSD)	never	–	0.896	5.5
Qwen3.8 27B	233	–	0.654	17.7
Qwen3.6 35B-A3B	228	140	0.659	100.2
Qwen3-Next 80B	222	34	0.776	71.4
Qwen 3.8 Flash Next (pruned)	77	2	0.335	40.3
GLM-4.7 Flash	10	7	0.205	83.7
Bonsai 2 9B	5	9	0.435	71.4
Gemma 4 26B	3	5	0.627	85.1
Qwen3.5 0.8B	2	2	0.094	349.8
DeepSeek-V2-Lite	1	1	0.225	136.1

• Every model that fit fully in 64 GB eventually crossed the same-meaning line. The best lasted about 230 rounds.
• The only two runs that held the meaning for all 1,000 rounds were the two big SSD-streamed MoE models (final similarity 0.91 and 0.90, final length 248–249 words vs 278). They ran about 5.5–5.8 tok/s, and each took about 17–19 hours of wall time.
• Heat (temperature 1.0 vs 0.7) made 4 of the 8 models collapse sooner, sometimes dramatically: Qwen3-Next 80B went from 222 to 34, Qwen3.6 35B-A3B from 228 to 140, and the pruned Qwen 3.8 Flash Next from 77 to 2. For Gemma 4 26B and Bonsai 2 9B, collapse came slightly later (3 → 5 and 5 → 9). It was unchanged for DeepSeek-V2-Lite and Qwen3.5 0.8B.
• Small models shrink the text fast. Qwen3.5 0.8B ended at 27 words with a final similarity of 0.094 (shown as +9.4% meaning in the video), and DeepSeek-V2-Lite ended at 13 words.
• Total compute: main 6.6 h, hot 2.4 h, frontier 35.9 h of generation wall time. In total, about 2.2M generated tokens.

Caveats

• One seed text, one prompt, one sampling seed per run. These are single trajectories, not averages. A different seed can move a collapse round a lot.
• The collapse rule and the 0.75 line are heuristics tied to MiniLM similarity. MiniLM is a small embedding model and measures topical and semantic closeness, not faithfulness.
• Quantization differs across models (2-bit to 4-bit, plus a pruned model), so this compares the deployed local checkpoints, not the base models.
• The frontier runs used OpenAI-compatible local servers (sushi / oMLX) instead of in-process mlx-lm. They were restarted/resumed across sessions, so their wall-clock span (Oct 23–25) is longer than their summed generation time. They report no peak memory, and the sushi build had MTP off. Their speeds reflect SSD streaming on this one machine.
• The Qwen3.8 27B hot run was not performed.
• No power or energy data was recorded.
• The "Qwen3.8", "Qwen 3.8 Flash Next", "GLM-5.3", and "DeepSeek V4.1" names are as published in the listed repos.
• Generated text is model output and may contain errors or odd phrasing. It is published as-is.

Files

• data/{main,hot,frontier}.parquet: per-round rows. data/summary.parquet: per-model-run summary.
• jsonl/: the same tables as JSON Lines.
• summary.csv: the summary without the long text fields.
• analyze.py: the metric/collapse code. build_dataset.py: builds this export from the raw run logs. seed_gettysburg.txt: the seed text.

License

Released under CC BY 4.0. The source text, the Gettysburg Address, is in the public domain. The model outputs were produced by the open-weight models listed above, and their own licenses apply to the models.

Citation

@misc{gandhi2026rewrites,
  author = {Pratik Gandhi},
  title  = {1,000 Rewrites Later: open-weight LLMs rewriting the Gettysburg Address 1,000 times on a Mac mini},
  year   = {2026},
  note   = {X: @pratikg}
}