Implementations
What you can actually run. Repository statistics verified September 2026.
The method-specific repositories are research code: self-contained academic releases at GPT-2 to 8B scale, without packaging and frequently pinned to a stale dependency snapshot. Most reproduce their paper's numbers. None is something to add to a serving stack.
The open weights are a different matter, and the gap between the two has widened. Ouro and Nanbeige4.2-3B both publish weights with model cards that say how to run them, and the latter was trained on 28T tokens and evaluated as a general agentic model and local assistant. Both repositories ship their own modelling code and both cards pin a Transformers version, and Nanbeige's other documented paths run through its own branches of SGLang, vLLM, llama.cpp and Ollama. Depth recurrence has reached checkpoints meant for use, even though the methods around it have not.
Open model checkpoints
Three sets of open weights reason by depth recurrence.
| Model | Scale | Notes | License |
|---|---|---|---|
| Ouro | 1.4B / 2.6B 7.7T tokens |
Looped language model, 24 layers, 4 recurrent steps by default with configurable early exit. Instruction-tuned -Thinking variants published alongside. Loads through Hugging Face Transformers with the repository's modelling code, pinned below version 4.56. The card notes that vLLM runs it without adaptive exit. Claims 3–4B dense-equivalent quality at 2.6B parameters. | Apache-2.0 |
| Nanbeige4.2-3B | 3B non-embed 28T tokens |
A general agentic model built on a looped transformer that reuses its layer stack for extra depth without extra parameters, trained on 28T tokens and released with open weights. Roughly thirty-five times the training tokens of Huginn and about four times Ouro. Its technical report evaluates it as a local assistant inside the OpenClaw agent framework. The card loads it with repository code under Transformers 4.45.1, and serves it through Nanbeige's own branches of SGLang, vLLM, llama.cpp and Ollama. An independent deployment report, run against Transformers 5.8.1 on Apple Silicon, documents five bugs in the released checkpoint, including a silently zeroed RoPE buffer and a model-construction error that is not specific to Apple hardware. | Apache-2.0 |
| Huginn-0125 | 3.5B 800B tokens |
Prelude / recurrent core / coda, with user-controlled test-time recursion
(num_steps, typically 4–64) and entropy, KL and argmax-stability
early-exit criteria. Eight intermediate training checkpoints also published. Explicitly
a proof of concept and undertrained relative to modern dense models of the same size. |
Apache-2.0 |
Method implementations
| Repository | Method | Stars / last push | State |
|---|---|---|---|
| facebookresearch/coconut | Coconut: continuous thoughts arXiv:2412.06769 |
1,705 · 2026-07 | Training code and data for the GPT-2-scale GSM8K / ProsQA / ProntoQA experiments. No official checkpoints; community reproductions exist on the Hub. MIT. |
| seal-rg/recurrent-pretraining | Recurrent depth (Huginn) arXiv:2502.05171 |
961 · 2025-12 | Full pretraining stack behind Huginn-0125. Well documented, actively maintained. Apache-2.0. |
| raymin0223/mixture_of_recursions | Mixture-of-Recursions arXiv:2507.10524 |
626 · 2025-09 | Complete pretraining framework with per-token depth routing and matched KV caching. The most production-adjacent of the architecture-level methods. Apache-2.0. |
| ezelikman/quiet-star | Quiet-STaR arXiv:2403.09629 |
740 · 2024-08 | Stale. Patches a specific transformers dev snapshot from 2024 into
Mistral's modeling code; brittle to run against current versions. Apache-2.0. |
| InternLM/SIM-CoT | Supervised implicit CoT arXiv:2509.20317 |
215 · 2026-04 | The best-resourced and most current implicit-CoT repository. Reports gains over CoT, Coconut and CODI on GPT-2 and Llama 1B/3B/8B. Apache-2.0. |
| da03/Internalize_CoT_Step_by_Step | Stepwise internalisation arXiv:2405.14838 |
208 · 2025-04 | Ships data (including GSM8K-Aug), pretrained checkpoints and eval scripts. Genuinely reproduces the paper. MIT. |
| zjunlp/LightThinker | Gist-token compression arXiv:2502.15589 |
165 · 2026-06 | Actively maintained. MIT. |
| zhenyi4/codi | CODI: self-distillation arXiv:2502.21074 |
107 · 2025-12 | Self-contained GPT-2-scale training and eval. Reproduces the GSM8K numbers. No license file: check before reuse. |
| xuyige/SoftCoT | SoftCoT and SoftCoT++ arXiv:2502.12134, 2505.11484 |
95 · 2025-05 | Training scripts, GSM8K preprocessing, checkpoints. NTU/NTUITIVE dual licence, non-commercial use only: commercial use, licensing, leasing and selling are expressly forbidden, and redistribution is restricted. |
| JacobPfau/fillerTokens | Filler tokens arXiv:2404.15758 |
76 · 2024-04 | Minimal research code for two synthetic algorithmic tasks. Dormant. No license. |
| mcleish7/retrofitting-recurrence | Retrofitted recurrence arXiv:2511.07384 |
69 · 2025-11 | Converts a pretrained non-recurrent model to depth-recurrent via curriculum. Apache-2.0. |
| yingfan-bot/lotus | LOTUS: supervised loop latents arXiv:2606.31779 |
51 · 2026-07 | Fine-tuning scripts over Llama-3.2-3B. no standalone checkpoint. MIT. |
| alperengozeten/CoT2 | Continuous CoT / parallel search arXiv:2505.23648 |
16 · 2026-02 | Accompanies the theory paper. MIT. |
| UW-Madison-Lee-Lab/looped-tf | Looped transformers, length generalisation arXiv:2409.15647 |
14 · 2025-04 | Small research codebase. |
Latent context and memory
Repositories for the work on the latent context page. Stars and last push read from the GitHub API on 9 October 2026.
| Repository | Method | Stars / last push | State |
|---|---|---|---|
| facebookresearch/context-language-models | Context Language Models | 637 / 2026-10-01 | Context as an editable file, zero-shot, in-context and RL training; Suffix Cache Reuse for serving. |
| microsoft/memento | Memento | 510 / 2026-04-08 | Block summaries with retained KV states, and the OpenMementos dataset. |
| Gen-Verse/LatentMAS | LatentMAS | 1,144 / 2026-06-18 | Training-free latent collaboration through a shared latent working memory. |
| thu-nics/C2C | Cache-to-Cache | 717 / 2026-10-09 | A trained fuser that projects one model's KV cache into another's. |
| HazyResearch/cartridges | Cartridges | 340 / 2026-03-23 | Compact KV caches trained per corpus by self-study. |
| bingreeky/MemGen | MemGen | 424 / 2026-09-24 | Latent memory tokens generated by a weaver, with an RL-trained trigger. |
| BytedTsinghua-SIA/MemAgent | MemAgent | 1,111 / 2026-05-12 | Overwritten fixed-size text memory trained with multi-conversation RL. |
| MIT-MI/MEM1 | MEM1 | 342 / 2026-01-03 | Constant-size internal state for multi-turn agents, trained with RL. |
| SakanaAI/doc-to-lora | Doc-to-LoRA | 832 / 2026-06-15 | A hypernetwork that writes a document into a LoRA adapter in one forward pass. |
| Continual-Intelligence/SEAL | SEAL | 1,870 / 2025-08-01 | Self-edits trained by RL against the updated model's accuracy. |
| test-time-training/e2e | TTT-E2E | 712 / 2026-02-15 | End-to-end test-time training in JAX, with checkpoints. |
Papers with no public implementation
- Pause tokens (Goyal et al., arXiv:2310.02226): never released by Google Research. The only related public repo is an independent toy that predates the paper and does not reproduce its pretraining setup.
- Compressed Chain of Thought (Cheng & Van Durme, arXiv:2412.13171): no release found anywhere. An unrelated multimodal-prompting repository shares the CCoT acronym.
- CoTFormer (arXiv:2310.10845): no official repository, only a later independent reproduction paper.
- Loop the Loopies! (arXiv:2607.16051): the largest looped model reported by total parameters, at 20B total and 2B active, so smaller than Huginn's 3.5B dense by active count. No public code or checkpoint.
- Neural Garbage Collection (Li et al., arXiv:2604.18002): the closest existing latent analogue of a context-managing language model, with no repository linked from the paper or found by search on 9 October 2026.
Benchmarks and data
Three instruments were built for this question, and they measure different objects.
| Instrument | What it measures |
|---|---|
| ProsQA arXiv:2412.06769 |
Search and planning on a synthetic DAG task, used to compare latent-CoT methods against each other. Roughly 17.9k train and 300 validation examples at 3–6 hops, shipped as JSON inside the Coconut repository rather than as an independent dataset. |
| Hagendorff & Fabi arXiv:2504.10615 |
Model-internal inference forced before the first answer token: the answer is signalled by the language of the initial response token, so it cannot be written out. 4,000 items, 18 models. The authors state they cannot rule out heuristic exploitation. |
| Think Fast arXiv:2606.07157 |
Breadth and scaling of no-CoT capacity: over 30,000 questions across 43 benchmarks, with no-CoT behaviour enforced, estimating a 50% task-completion time horizon against human times. A suite rather than a dataset, and its collated datasets are available on request rather than released. |
Everything else is borrowed: ProntoQA, GSM8K-Aug (385k augmented samples with intermediate expressions stripped, now the de facto training corpus for implicit CoT), GSM-Hard, and the standard suites MATH-500, AIME, SVAMP, BIG-Bench Hard, CommonsenseQA, ARC-C and GPQA.
What is still missing is a shared instrument that discriminates between latent-reasoning architectures and rules out shortcut exploitation. The three above measure different things, none has been adopted as a community standard, and critique papers continue to construct bespoke diagnostics rather than reach for a common one. Cross-paper comparison remains correspondingly unreliable.
Framework support
No cross-framework API was found for generic continuous-thought decoding, pause tokens,
user-controlled recurrent depth, or inspection of intermediate loop states. Huginn and Ouro
ship custom modeling files loaded through trust_remote_code=True, a packaging convenience rather than a library feature.
One architecture did get native support, and its lifecycle is more informative than the
absence would be. vLLM merged
OuroForCausalLM on
30 October 2025, a 521-line implementation containing the recurrence loop itself rather than
deferring to the repository's custom code, with Ouro-1.4B and 2.6B listed in the
supported-model table. vLLM
removed it on 25 July 2026,
giving three reasons: the model was about a year old and superseded by Plamo3, it saw “very little
usage in vLLM,” and it had been marked for deletion.
A major serving engine integrated a looped architecture and then withdrew it for lack of demand. That is a sharper reading of ecosystem maturity than never having been integrated at all. What the integration was not also matters: inference support for one architecture, not a general interface for latent reasoning.
Non-generative classifiers
An adjacent family decides without generating text, and is packaged as products rather than research code. Knowledgator/GLiClass (523 stars, last push 2026-07) and fastino-ai/GLiNER2 (2,132 stars, 2026-09) are pip-installable libraries under Apache-2.0, with Apache-2.0 checkpoints on Hugging Face. The GLiNER2 repository's examples load GLiNER2.5 boundary checkpoints, a later design than the published paper describes, so a result should name its checkpoint. Jev is a hosted API in early access, and its launch post describes the architecture only as new and names no downloadable weights. The routing page places these models among text-based routers.
Curated lists
-
Awesome-Latent-Space —
970 stars, last updated July 2026. The broadest and most-starred. (Formerly published under
the
YU-deepaccount. That path now redirects.) - EIT-NLP/Awesome-Latent-CoT — 382 stars, June 2026. Companion to the "Reasoning Beyond Language" survey and the most focused on latent chain of thought specifically.
- LatentCoT-Horizon — 414 stars, November 2025.