
Comparing 36 'next-generation Transformers' on one yardstick: an open-source leaderboard from a first-year PhD student at HKU, 9 days, 390 stars
What it is: an open-source training framework plus a public leaderboard. Dozens of new LLM architectures are trained on the same data with the same compute to see which really performs better, and AI agents can also submit new architectures to the board.
Why it took off: it rode the 'next-gen architecture' buzz of 2026 autumn. On launch day it shipped 'who won' results from several hundred GPU-days of runs, then kept people returning with 'unlock 300 days of updates at 28 stars', pushed simultaneously by the author and team members on Xiaohongshu, X, and WeChat groups.
CHow you could follow Don't copy the LLM architecture route. Pick a much cheaper, currently hot direction where papers contradict each other (agent frameworks, RAG, quantization, samplers). Finish a batch of same-budget results before open-sourcing, lead with a counterintuitive conclusion in the first line, and have several people repost on launch day.
- Data (updated daily) · Repo created 2026-10-31 · 483 stars · 7 days +75
GitHub · stars Data 2026-11-02 → 2026-11-20 (19 days, 19 observations). Points before 10-28 come from deeply / archive backfill, not every day has one
A first-year PhD student at HKU MMLab recruited 7 doctoral students and trained 36 model architectures across six families (attention, sparse, linear, recurrent, residual, hybrid) on the same data and training compute, then published the results as a website and open-source leaderboard. The hottest 'recurrent Transformers' all lost to a plain baseline on this board. Backed by real compute, good timing, and a Chinese academic-community launch playbook, it gathered about 390 stars in 9 days. Below: why it took off, where the stars came from, and what you can and can't copy if you want a project that earns stars the same way.
A Nov 1 noon post
On 2026 Nov 1 Beijing time 13:14, Yue Su (Yue Su on X as @Selen7005717917, bio lists HKU MMLab, ByteDance Seed, SJTU MVIG) posted a long thread, opening with:
Which Transformer architecture will win in the next era?
Five points below: 36 models share one training harness; all models use the same data, same tokens, same optimizer, with training compute matched to within 1% and fully audited; 31 runs' results can be replayed; on one 8-GPU machine you can train a model of about 1.4B parameters in 2–4 days; the last one: "Hand an Arena task to any LLM, and the architecture it proposes is trained under the same rules and posted on the same leaderboard." The thread ends with "Links in the replies, QR code on the first image."
As of 11-08 this thread had only 2,136 views and 25 likes (X, checked via fxtwitter). The real launch happened elsewhere: 42 minutes later, Yue Su posted two notes on Xiaohongshu (account "selen", 1,836 followers, more than double his X following). Two teammates, "风里风" and "Ä", reposted at the same time. Within two hours there were 4 notes, and the most shared title line was "36 llm architectures, finally a fair comparison."
The repo was created on 10-31. As of 11-09 it had 393 stars and 31 forks.
What it actually is
An open-source training framework plus a public leaderboard (MIT license). The pain point is the first line of the README: When the training budget is fixed, how should a language model spend its compute? Each architecture paper uses its own data, steps and evals, so you can't tell whether a result comes from the architecture or from more training. NextTransformer locks every variable: the same data (karpathy's ClimbMix), the same 49.152B tokens, matched theoretical training compute per token (about 4.67 GFLOPs), and evaluation on 22 CORE benchmarks at the same checkpoints. Parameter count, memory and speed are outputs, not controls. The base is karpathy's nanochat, and linear attention operators come straight from flash-linear-attention.
The scorecard is the most shareable part. On the 11-08 leaderboard, 36 runs. A few rows (higher CORE is better, compared against the full-attention baseline):
| Architecture | Category | CORE | vs baseline |
|---|---|---|---|
| AttnRes-Canon | Residual | 35.99 | +2.90 |
| mHC | Residual | 35.33 | +2.23 |
| Hybrid GDN + SWA | Mixed | 33.66 | +0.57 |
| Full-attention baseline | Baseline | 33.10 | — |
| Fixed Recursive (best of the loop variants) | Loop | 32.45 | −0.65 |
| Gated DeltaNet | Linear | 30.55 | −2.55 |
| Ouro | Loop | 27.70 | −5.40 |
(project leaderboard page, 11-08)
All 8 looped architectures ranked below the baseline. The author wrote a candid line under the table: each config was run with one seed, on different machines, and "the ranking is an observation, not a statistical conclusion." For readers, though, "at equal compute, the looped models everyone has been hyping didn't win" is already a conclusion worth sharing.
The other half is Arena: give an agent a task brief, let it propose new architectures itself, train them under the same rules, and post them on the same leaderboard. The full Arena v2 is promised for 11-14.
Why now
In autumn 2026, "next-generation architectures" were the hottest topic in the LLM world. After September's GPT-6 Astra release, Sebastian Raschka 9-09 wrote a long piece on looped transformers; 11-02 SenseTime open-sourced Loop-9B; hybrid linear attention (GDN, KDA, the Qwen3-Next line) has become a regular in every lab's tech report. Lots of buzzwords, but nobody had measured them with the same ruler. NextTransformer gathered Ouro, AttnRes, mHC, GDN, KDA, NSA and MoBA in one place, and on launch day it already gave a verdict on who wins.
It also rides on karpathy's ecosystem: the nanochat base, the ClimbMix data, and an Arena concept that maps onto autoresearch (100k stars). Anyone who has followed those two projects gets it immediately.
Where the stars came from
GitHub stars per day:
| Date | star | Added that day |
|---|---|---|
| 11-02 | 89 | — |
| 11-03 | 165 | +76 |
| 11-04 | 200 | +35 |
| 11-05 | 275 | +75 |
| 11-06 | 315 | +40 |
| 11-07 | 346 | +31 |
| 11-08 | 374 | +28 |
| 11-09 | 393 | +19 |
There was no single spike. The first few days brought a few dozen a day, then it trailed off steadily. The traffic sources I could see are these:
- Xiaohongshu is the home turf. On launch day, the 4 notes together got about 250 likes. The author kept posting afterwards: 11-05 "400B tokens pretraining, join us in choosing the plan" (66 likes, 23 comments, the most-commented one), 11-06 "28-day update schedule", a daily Day 1 and Day 2 post. Yue Su had already run "HKU PhD doing AI" on Xiaohongshu, and his notes on the previous project also got over a hundred likes.
- X traffic is small. The launch thread got 2,136 views. A reply from 11-03, "Added DeltaNet + GDN today, Discord opens tonight, 300 stars means the 28-day update schedule starts," got 1,215 views. I found no big-account reposts.
- WeChat groups and Discord. The QR codes appear on the thread's first image, the Xiaohongshu posters and the README. How far they spread inside groups isn't visible from outside.
- Each of the 7 team members reposting from their own lab (inferred). Contributors are spread across HKU MMLab, Tsinghua, CASIA, the University of Tokyo and SJTU. Early forkers cluster at SJTU, Zhejiang University and ICT, CAS. Each person reposting in their own group fits the curve of "dozens a day, no spike."
- Aggregator sites that auto-scrape new GitHub repos: xiaodu.blog (11-01), the Xiangshang Research daily at #3 (11-03, editor's words: "solves the chronic problem that current architecture papers can't be compared fairly"), AiBoss (11-04), and others, long tail.
Not present: Hacker News, Reddit, Zhihu, Bilibili, WeChat official accounts like QbitAI or Synced, and GitHub Trending (repos under 60 days old can't make it). So the 390 stars came almost entirely from a small Chinese academic circle passing it along.
Since launch it has kept giving people reasons to come back: 300 stars in 28 days of updates, 11-14 Arena v2, and a public call for 200B / 400B token proposals (GitHub Discussions #22). Outside contributors have started adding to it too: an Italian developer proposed Gated Attention, someone built a multimodal branch called NextTransformer-V, and a team at Peking University's tooling sped up baseline training by 1.99x. All of these are featured in the README's "Community Projects" card to give contributors some spotlight.
Who else does the same thing
| Project | Scale (11-08) | How it differs |
|---|---|---|
| karpathy/autoresearch | 99.6k stars | Lets an agent do automated research on a single GPU, any direction, no compute matching |
| karpathy/nanochat | 60.4k stars | Its base, for end-to-end small models, not architecture comparison |
| KellerJordan/modded-nanogpt | 6,000 star | Wall-clock time for a 124M-parameter model to reach a target loss; a speedrun contest |
| openai/parameter-golf | 5,200 star | Compete on a 16MB model size limit |
| fla-org/flash-linear-attention | 6,000 star | Linear attention operators and a training framework, no public leaderboard |
| InternLM/archspace | 89 star | The closest architecture-exploration library in direction, but no large-scale results under the same protocol |
(GitHub,11-08)
Speedruns rank by time, size contests rank by size, operator libraries give parts. "Fixed compute, dozens of new architectures, fully audited, a leaderboard with a website" has no direct competitor yet, and I haven't found anyone copying it.
What can be copied and what can't
The results can't be copied. Each config is 1.4B parameters and 49B tokens. On one 8-GPU machine that's 2–4 days; across 31 configs it works out to roughly 500–1,000 GPU-days (estimate), and the data needs 600GB to download first. The author didn't say where the compute came from (ByteDance Seed internship, the HKU cluster, or somewhere else, unverified). The format can be built in a week, but a leaderboard without results doesn't get shared.
Three things can be copied:
- The format: "same-conditions leaderboard + model zoo + replayable website + an agent that climbs the board." Any field where "papers talk past each other" can take it: RL algorithms, diffusion samplers, recommendation models, agent frameworks, RAG schemes, quantization methods. In many fields a run takes a few GPU-hours on one card.
- Have a conclusion on launch day. Finish the runs before publishing, and lead with "X lost, Y won," not "everyone welcome to run it."
- The launch playbook of the Chinese academic scene: the main account plus teammates posting on Xiaohongshu at the same time, posters with QR codes to pull people into groups, "this many stars unlocks the update plan," a public call for the next phase's plans, and outside contributions written into the README.
It also has a ceiling: no big accounts, no English community relay, so after 9 days and 390 stars it's only gaining a dozen or twenty a day. Whether 11-14's Arena v2 and 400B token plan ship on schedule will decide whether this is a one-wave launch spike or a benchmark that keeps getting cited long-term.
If it were you
Verdict: C. You can learn from it, but copying it isn't worth the cost. Following it in its own field means matching hundreds of GPU-days and a lab's backing, which latecomers can't get. Building a "same-ruler" leaderboard in a different field is doable, but projects of this kind top out at a few hundred stars. This one had the timing, the team and the compute, and it only reached 9 days and 390 stars.
Who it fits: PhD students or people in a lab, holding a research direction where nobody can say "who's best," with compute or cluster quota in hand, and a small circle (lab groups, Xiaohongshu, X) that will repost together on launch day.
Who it doesn't fit: solo developers with no compute who want to copy a leaderboard for LLM architectures; people with no launch channels who expect organic GitHub traffic (repos under 60 days old can't reach Trending, and stars depend on first-week reposts).
How to do it: Pick a direction that costs about two orders of magnitude less than LLM pretraining but is hot right now (agent frameworks, RAG, small-model quantization, RL post-training recipes, diffusion samplers). Set a fixed-budget protocol, run 15–30 experiments yourself first, and publish with a clear "who won, who lost" conclusion and a site where people can replay the curves before you open-source it. Leave an entry point where agents propose their own setups and compete on a leaderboard under the same rules, so you ride the autoresearch narrative.
How to get users: lead with a counterintuitive conclusion ("at the same budget, the hottest X didn't win"). For the Chinese community, go through Xiaohongshu: the main account plus 3–4 collaborators posting at the same time, with posters carrying QR codes to pull people into groups. For the English community, add a step: get one or two researchers with followings in the field to preview the results and repost on launch day. After launch, drop a new set of results every two or three days so the curve doesn't fall off by day three.
Deadline: the buzz around "next-gen architecture" peaks in 2026 fall. If you're building in a different field, ship during the hottest one or two months of the area you pick. On the LLM architecture track, it gets harder to catch up after 11-14 Arena v2.
Where the call could be wrong: ① After 11-14, Arena v2 ships on time and daily stars climb back above 50, or a big account like karpathy reposts it. That would show the "same ruler" alone can earn thousands of stars, so the rating should go up. ② The reverse: within two weeks daily additions drop to single digits and the 28-day update plan stalls. That would mean it was a one-off launch spike and the format itself lacks lasting pull. ③ A big company or Hugging Face publishes a public architecture leaderboard on the same basis. Then the position is taken.
Research notes: competitor table, acquisition evidence, full research report
- Competitors
Name Where Launched Scale How it differs from this one karpathy/autoresearch Open source 2026-03-06 99.6k star General-purpose agent automated research, single-card nanochat, no compute matching. NextTransformer only does architecture, same FLOPs and same tokens for the leaderboard. karpathy/nanochat Open source 2025-10-13 60.4k star The base for NextTransformer. Does end-to-end small models, not architecture comparisons. KellerJordan/modded-nanogpt Open source 2024-06-01 6.0k star Wall-clock time for a 124M model to reach a target loss. NextTransformer fixes theoretical FLOPs, runs at 1.4B / 49B scale, and compares by architecture family. openai/parameter-golf Open source 2026-02-09 5.2k star A contest with a 16MB model size limit. Its constraint dimension is different. fla-org/flash-linear-attention + flame Open source 2023-12 / 2025-01 6.0k / 426 star Linear attention kernels and a training framework (NextTransformer uses its kernels directly). No unified protocol and no public leaderboard. InternLM/archspace Open source 2026-07-21 89 star The closest architecture-exploration knowledge base in direction, but no large-scale training results under a shared protocol. HazyResearch/zoology / erogol/BlaGPT / google-deepmind/simply Open source 2023–2025 291 / 118 / 590 star A testbed or research codebase for architecture experiments on synthetic tasks or small data. No compute audit and no website-based leaderboard. yifanzhang-pro/recurrent-looped-tranformer, sandyresearch/parcae Open source 2026-09 / 2026-02 1,141 / 191 star Implementations of a single looped method. NextTransformer takes methods like these in as contestants and compares them. Internal ablations at each lab (Qwen3-Next, Kimi AttnRes, SenseTime Loop-9B, etc.) Big-company features 2025–2026 Not public Big labs run the same kind of comparison internally, but they don't publish results under the same protocol - Where users come from
- Verified Xiaohongshu (RedNote): 11-01 05:56–07:54 UTC, author selen and team members 风里风 and Ä posted 4 notes in a row (81/113/20/33 likes); then an RFC post with 66 likes and 23 comments, and a 28-day update plan, Day 1/2; the author has 1,836 followers (TikHub)
- Verified X: 11-01 author's long-form first post 2,136 views 25 likes 13 bookmarks, first image has a QR code; 11-03 reply "300 star unlocks 28-day update plan, Discord tonight" 1,215 views (fxtwitter); no big-account reposts found
- Verified WeChat group / Discord: the README's community section links a permanent Discord invite and a WeChat group QR code, and the poster carries a QR code. A WeChat Official Account search turned up no related articles.
- Inferred Each of the seven team members' lab circles: contributors come from HKU MMLab, Tsinghua, CASIA, the University of Tokyo, and SJTU. Early forks cluster at SJTU, Zhejiang University, and ICT-CAS. The star curve has no single spike, just 4 consecutive days of 50–80.
- Verified auto-aggregator sites: xiaodu.blog 11-01, 向上研究所 11-03 daily report #3, AiBoss 11-04, AISignal 11-06, Kapyn, Vigía
- Verified Riding the karpathy ecosystem: the acknowledgments say it builds on nanochat and uses karpathy/climbmix-400b-shuffle; the Arena maps to the autoresearch concept.
Interactive-Active-AI/NextTransformer: 36 "next-generation Transformer" architectures benchmarked head-to-head at the same compute
Research date 2026-11-08. Repo github.com/Interactive-Active-AI/NextTransformer, observatory interactive-active-ai.github.io/NextTransformer/, MIT license.
Bottom line
NextTransformer is an open-source framework, benchmark and leaderboard built by Yue Su (Selen on X @Selen7005717917, "selen" on RedNote), a first-year PhD student at HKU MMLab, with 7 PhD students from HKU, Tsinghua, CASIA, the University of Tokyo and SJTU. It trains six families of architectures (attention, sparse, linear, looped, residual, hybrid), 36 variants in total, on the same data, with 49.152B tokens and training compute matched to within ±1%, and has published results from 31 completed runs. It hit about 390 stars in 9 days, riding three things at once: a topic that matched the architecture hype of fall 2026 (the looped transformer popularized by GPT-6 Astra, SenseTime's Loop-9B, and everyone's hybrid linear attention); results from real compute (1.4B-parameter models, 8 GPUs for 2–4 days each, roughly 500–1,000 GPU-days across 31 runs, the hardest part for a small team to copy); and a launch playbook straight from Chinese academic circles: RedNote + X + WeChat groups (the author has 1,836 RedNote followers, and on launch day the author and two teammates posted 4 notes between them with QR-code posters, then kept the buzz going with a "300 stars unlock 28 days of updates" plan). It's not big yet: the X launch post got only 2,136 views, it didn't reach HN, Reddit, or the top 25 on Trendshift, and no major accounts or media covered it. It's circulating in a small circle for now.
1. Data
- GitHub stars (our repo_day / data card): repo created 2026-10-31 07:46 UTC; 11-02 first observation 89 → 11-03 165 → 11-04 200 → 11-05 275 → 11-06 315 → 11-07 346 → 11-08 374 → 11-09 393. Up 304 over 7 days, single-day peak 76. 31 forks, fork/star ratio about 8%,
suspectis empty. - Trendshift daily star gains (https://trendshift.io/repositories/330637): 11-01 +78, 11-02 +67, 11-03 +48, 11-04 +70, 11-05 +36, 11-06 +28, 11-07 +31, 11-08 +15. 10 days with merged PRs: 22. Not in Trendshift's top 25; the repo is under 60 days old, and GitHub trending doesn't pick up repos that young (an empirical pattern from our own repos). Growth has clearly slowed.
- Contributors (GitHub API): yitianlian (谢承兴) 25, Selen-Suyue (苏越; bio lists ByteDance-Seed / HKU-MMLab / MVIG-SJTU) 24, ggeeyyi (Tsinghua) 13, Xiangnan-Wu (CASIA) 5, Xiangyue-Zhang (University of Tokyo, interning at Tencent) 4, xjxyys (HKU MMLab) 3, henry-y 3. Most of the 22 merged PRs were submitted by the core team. External contributions: matteodisalvo (Italy) proposed the Gated Attention method (issue #12 / PR #13); SnifferCaptain proposed YOCO (#19); Enderfga built the multimodal branch NextTransformer-V (#14); henry-y and PKU's DAGroup contributed ResearchBoost-Agent for training speedups (dense_full 8× on 1.99 GPUs, PR #26); drewjin (SJTU) filed issue #8.
- Who forked it (GitHub API): the first few forks are team members and students at SJTU, Zhejiang University, HKU, and ICT-CAS. Another dozen or so are brand-new 0-follower accounts. Nothing looks like star-botting; it reads more like word of mouth in academic circles.
- Stargazer breakdown: the GitHub endpoint returned 404 / required auth this time, so we couldn't pull it and can't profile stargazers by source (unverified).
- Organization: Interactive-Active-AI registered 2026-08-02, with 3 public repos: NextTransformer 394, OneWM 92 (2026-10-15, the author's world-model paper, arXiv 2610.21095), and Aha-Looped-Transformer 38 (2026-09-15, a looped-transformer map). A same-named organization on HF, NextTransformer, has 6 members and 0 models.
2. Competitors (doing the same thing)
| Name | Where | Launched | Scale | How it differs from NextTransformer |
|---|---|---|---|---|
| karpathy/autoresearch | Open source | 2026-03-06 | 99.6k star | Lets an agent do automated research on a single-GPU nanochat, with no fixed direction and no compute alignment. NextTransformer's Arena focuses purely on architecture and judges winners at the same FLOPs and 49B tokens. |
| karpathy/nanochat | Open source | 2025-10-13 | 60.4k star | NextTransformer's foundation (the acknowledgments say it builds on nanochat and karpathy's ClimbMix). nanochat is end-to-end "100-dollar ChatGPT" and doesn't do architecture comparison. |
| KellerJordan/modded-nanogpt | Open source | 2024-06-01 | 6.0k star | Speedrun leaderboard: wall-clock time for 124M models to reach a target loss. NextTransformer fixes compute and compares quality, scales to 1.4B / 49B, and organizes results by architecture family. |
| openai/parameter-golf | Open-source competition | 2026-02-09 | 5.2k star | Limits model size (16MB) and compares quality; the constraint is different |
| fla-org/flash-linear-attention + fla-org/flame | Open source | 2023-12 / 2025-01 | 6.0k / 426 star | Linear-attention kernels and training frameworks. NextTransformer uses FLA's kernels directly. FLA doesn't publish a public leaderboard at matched compute. |
| InternLM/archspace | Open source (Shanghai AI Lab) | 2026-07-21 | 89 star | "Turning architecture exploration into reusable knowledge"; closest in direction, but no large-scale, same-protocol results to compare |
| HazyResearch/zoology, erogol/BlaGPT, google-deepmind/simply | Open source | 2023–2025 | 291 / 118 / 590 star | Architecture testbeds on synthetic tasks or small datasets / research codebases; no unified compute audit and no web leaderboard |
| yifanzhang-pro/recurrent-looped-tranformer, sandyresearch/parcae | Open source | 2026-09 / 2026-02 | 1,141 / 191 star | Official implementations of a single method (looped). NextTransformer takes methods like these in as competitors and compares them. |
| Paper side: SMELT (arXiv 2609.01343, scaling laws for MoE looped models at matched FLOPs), plus ablations in various labs' technical reports | Paper | 2026-09 and others | — | Each paper uses its own configuration, so there's no apples-to-apples comparison. That's exactly the pain point the first line of the NextTransformer README describes. |
| Big-company features | — | — | No matching product | Qwen3-Next, Kimi AttnRes, SenseTime Loop-9B and others all run similar comparisons internally, but none publish results under the same protocol. |
Bottom line: "same compute, same data, publicly audited, covering dozens of new architectures, plus a leaderboard and website" has no direct competitor right now. The nearest are either speedrun / size-limit contests (modded-nanogpt, parameter-golf), kernel / training frameworks (FLA), or general auto-research (autoresearch, 100k stars, where the buzz lives).
3. Growth tactics (where users come from)
- Xiaohongshu (confirmed, the main Chinese channel). Searching "NextTransformer" on TikHub: 11-01 05:56–07:54 UTC, 4 posts in two hours. The author "selen" posted two (81 likes 43 saves, and 113 likes 68 saves); teammate "风里风" posted one (20 likes, "We open-sourced NextTransformer"); teammate "Ä" posted one (33 likes 19 saves, "36 LLM architectures, finally a fair comparison"). The author kept posting: 11-05 "400B-token pretraining, join us in shaping the plan" RFC (66 likes, 23 comments, the most-commented post), 11-06 "28-day update plan," 11-06 Day 1 multimodal, and 11-07 Day 2. Community member Henry posted "RBA×NextTransformer" (15 likes, 11-07). The author has 1,836 Xiaohongshu followers, more than double the X count. Earlier posts (GPT6 controlling a real robot, 112 likes; OneWM, 101 likes) show he has long built a persona as "an AI PhD from HKU."
- X (confirmed, but small reach). The launch long-post went up at 11-01 05:14 UTC (https://x.com/Selen7005717917/status/2116760739826565464; per the fxtwitter API: 2,136 views, 25 likes, 5 reposts, 13 bookmarks), with a QR code on the first image. A reply said "300 stars opens the 28-day update plan, Discord goes up tonight" (11-03, 1,215 views). The author has 798 X followers. Trendshift's "Social mentions" only picked up these 1 posts. No big-account reposts found (the local X corpus x-radar returns no mentions, and a Brave search of x.com found nothing either).
- WeChat groups / Discord (confirmed to exist, reach unknown). The README's community section includes a permanent Discord link and a WeChat group QR code page; the X first image and the Xiaohongshu poster both carry QR codes. WeChat search (TikHub) turned up no Official Account articles.
- Team members' academic circles (inferred, indirect evidence). The 7 contributors are spread across HKU MMLab, Tsinghua, CASIA, the University of Tokyo, and SJTU MVIG, and Yue Su is also listed at ByteDance Seed. Early forkers cluster at SJTU, Zhejiang University, and ICT-CAS. Xiangyue Zhang lists it in the publications on his personal homepage. This kind of "each person shares it in their own lab group" diffusion doesn't show up directly in the data, but it fits the star curve: no single spike, just 4 consecutive days of 50–80 each day.
- Aggregator sites (confirmed, long tail). xiaodu.blog (11-01, Chinese-language write-up), 向上研究所 amazingindex (#3 on the 11-03 daily list, recommendation score 75), AiBoss (11-04, in Chinese, English and Japanese), AISignal (11-06; the content is clearly AI-generated and the metrics look made up), Kapyn, Vigía. All of them scrape new GitHub repos and auto-generate articles, so the traffic they bring is limited.
- Not present: Hacker News (our hackernews table has no match), Reddit, Zhihu columns, Bilibili, WeChat Official Account coverage from 量子位 / 机器之心 / 新智元, Trendshift, paid promotion.
4. Why it took off
- Hits a real pain point: architecture papers each use their own data, step counts and evals, so "is the architecture better, or was it just trained longer?" can't be settled. The README's first line is "Under a fixed training budget, how should a model spend its compute?" The 向上研究所 editor's verdict: "it solves the chronic problem that current architecture papers can't be compared fairly."
- Timing: autumn 2026 is when "next-gen architecture" was hottest. September: GPT-6 Astra launched, Raschka wrote a long piece on looped transformers (2026-09-09); 11-02 SenseTime open-sourced SenseNova Loop-9B; Zhipu put Loop Transformer on its roadmap; linear and hybrid attention (GDN, KDA, Qwen3-Next) are now mainstream talk. Zhihu Frontier's June piece was already arguing the next Transformer will absorb looping and sparse routing. NextTransformer pulls all these buzzwords (Ouro, AttnRes, mHC, GDN, KDA, NSA, MoBA) into one place and gives an answer to "who won": at equal compute, mHC, AttnRes and the GDN hybrid all beat the dense baseline on BPB and CORE.
- Real results, not just code: 31 groups trained on 1.4B / 49B tokens, full learning curves per group and CORE-22 across 6 checkpoints, all presented as a replayable website. At 8 GPUs for 2–4 days per group, 31 groups come to about 500–1,000 GPU-days (1.2–24k GPU-hours). Where the compute came from is not disclosed (unverified). Copying the format is easy; copying the results is hard.
- Rides the karpathy ecosystem: built on nanochat, uses karpathy's ClimbMix data, and the Arena concept maps directly onto autoresearch (100k stars). Anyone who has seen autoresearch gets it immediately: "let an agent design the architecture and rank it under the same rules."
- Packaging and ongoing updates: bilingual Chinese/English README, movie-poster-style cards, an observatory website, model cards and architecture diagrams. After launch it gives people reasons to come back: "300 stars unlocks a 28-day update plan", "11-14 launches Arena v2", "400B-token RFC for proposals". It also puts external contributions (NextTransformer-V, ResearchBoost-Agent) in a "community projects" card in the README, giving contributors exposure.
5. Differentiation
Compared with speedrun projects (modded-nanogpt), it fixes theoretical training FLOPs rather than wall-clock time or parameter count, treating "parameters, memory and speed are outcomes, not controls." Compared with kernel libraries like FLA, it adds a unified protocol, auditing and a public leaderboard. Compared with autoresearch, it confines agent research to architecture alone and scores everything on one yardstick. That differentiation is why people share it: "36 architectures can finally be compared fairly" is itself a shareable line. But the scale it can currently reach (a few hundred stars) is mostly limited by a narrow launch channel and a high reproduction barrier (8 GPUs for 2–4 days, 600GB of data). So the differentiation explains why people share it, but it hasn't turned into mass spread yet.
6. Copycats
None found. Other repos named NextTransformer are forks. NextTransformer-V and ResearchBoost-Agent are extensions built on top of it, not copies.
7. What others can learn
- The format of "same-condition leaderboard + model zoo + observable website + agent-run leaderboard" can be transplanted to any field where papers talk past each other (RL algorithms, diffusion sampling, recommendation models, agent harnesses).
- Run a batch of real results on real compute before open-sourcing, so there is a "who won" conclusion to spread on the day of release.
- The Chinese academic-circle launch playbook: main account on Xiaohongshu plus team members posting simultaneously, QR-code posters, WeChat groups / Discord, and "X stars unlocks an update plan" plus a public RFC, so something new arrives every day for a week after launch.
- Writing external contributions into README card sections, rewarding contributors.
8. Uncertain
- The source of the stars can't be broken down: the stargazer endpoint doesn't expose it, and how much got shared in WeChat groups isn't visible.
- Where the training compute came from (ByteDance Seed internship, HKU MMLab cluster, or elsewhere) hasn't been disclosed.
- Whether the second small spike in 11-04 (+70) corresponds to the Discord launch, the AiBoss write-up, or some WeChat group can't be determined.
- Whether 11-14 Arena v2 and the November 400B token plan ship on time will decide if it's a one-off launch spike or becomes a lasting community benchmark.
Sources
- GitHub API: repos, contributors, PRs / issues, fork lists (2026-11-08 lookups)
- Our github tables repo_profile / repo_day; hackernews table (no results)
- Trendshift:trendshift.io/repositories/330637
- X (fxtwitter API): x.com/Selen7005717917/status/2116760739826565464, x.com/Selen7005717917/status/2117509847876833577
- Xiaohongshu (TikHub app_v2/search_notes, get_user_info, get_user_posted_notes; user 62d3ed3c0000000002000fed)
- Aggregators: xiaodu.blog/articles/390, www.amazingindex.com/article/00a8ed983a7387309d4bacce9072733…, www.aiboss88.com/en/news/project-nexttransformer, www.aisignal.dev/analysis/interactive-active-ai-nexttransfor…
- Timing: magazine.sebastianraschka.com/p/gpt-6-astra-looped-transform… (2026-09-09), www.ithome.com/1/020/973.htm (SenseTime Loop-9B, 2026-11-02), x.com/ZhihuFrontier/status/2070442689427058900
- Competitor stars: GitHub API (2026-11-08)