Organization repository on GitHub
NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness.
- Forks
- 21
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Models & Reimplementations
NeoHorse
Open-weight 4B and 9B models from TokenRhythm: NeoHorse-1 for tool use, coding, and instruction following, plus NeoHorse-Jev-4B for prefill-only Choice, Noul, and Score decisions.
- Stars
- 1.7k
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- TokenRhythm
- Last verified
- Oct 2026
About this project
NeoHorse Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness. NeoHorse is TokenRhythm's family of open-weight models for agent workflows. NeoHorse-1 provides 4B and 9B language models for tool use, coding, and instruction following. NeoHorse-Jev builds on NeoHorse-1-4B to make structured decisions through prefill-only inference. [2026-09-24] 🚀 NeoHorse-Jev-4B is here! Our new decision model brings Choice, Noul, and Score to the NeoHorse family. Explore the release.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- NeoHorse-Jev-4B uses prefill-only inference to return a Choice, a Noul, or a Score with probabilities.
- Reports 77.70 across six text benchmark groups and 83.26% mean accuracy on Nimble, VitaminC, and MASSIVE.
- Weights ship as BF16 plus 8-bit, 5-bit, and 4-bit GGUF, with MLX builds for Apple silicon.
- NeoHorse-1 checkpoints are post-trained from Qwen3.5 and released under Apache 2.0.
Watch out
The benchmark numbers are the project's own evaluation, and the comparison table only ranks models that reported all six groups, so read the ranking as a maintainer claim rather than an independent result. Nothing documents multimodal input, serving configuration, throughput, or latency; the README only says to download the complete bundle and use the matching runtime or backend adapter. The captured README is cut off before the full model details table, so per-checkpoint specifications are not visible. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
Not as documented here. Applications define the questions and possible answers, then read structured decisions and probabilities to route requests, select tools, or control a workflow, and the README tells you to download the complete model bundle and use the matching runtime or backend adapter. It does not enumerate those adapters or give a serving command, so you would be reading the model card for that.
By the project's own table it averages 77.70 across six text benchmark groups, against 77.13 for Open-Jev-9B and 74.25 for Kev-4B. It leads on JevBench, Kev, Nimble, and VitaminC, and trails on OpenJev text and MASSIVE. No independent rerun of these numbers is cited, and only models with all six groups are ranked at all.
The README lists BF16 weights alongside 8-bit, 5-bit, and 4-bit GGUF builds plus MLX versions for Apple silicon, and presents the 4B size as the lighter deployment footprint. It gives no memory figures, no throughput numbers, and no comparison of accuracy lost at each quantization level, so you would be choosing a build without measurements.
As described, it assigns tasks to a heterogeneous model pool, records tool interactions and outcomes, estimates capability demand, and feeds capability-level feedback into the next training mixture, forming an evaluation-selection-update loop the README calls a prototype. It is presented as research infrastructure on a path toward recursive self-improvement, and no public harness package, interface, or install step is described.
NeoHorse is TokenRhythm's open-weight family for agent workflows. NeoHorse-1 provides 4B and 9B language models for tool use, coding and instruction following, and NeoHorse-Jev builds on NeoHorse-1-4B to make structured decisions through prefill-only inference.
It scored 77.70 across six text benchmark groups, the highest among the four open-weight decision models with complete results, and 83.26% mean accuracy across Nimble, VitaminC and MASSIVE, which is 11.50 percentage points above NeoHorse-1-4B.
The models are released under Apache 2.0, with 4B and 9B GGUF and quantized versions on Hugging Face including BF16, 8-bit, 5-bit and 4-bit weights, plus MLX versions for Apple silicon. NeoHorse-Jev-4B returns structured decisions and needs its complete model bundle plus the matching runtime or backend adapter.
More Jev resources like this
1.5kGitHub stars
A 0.8B decision model that answers first and escalates only when unsure, plus nine LoRA adapters of roughly 41 MB each for guard, triage, tool choice, spam, and similar classification work.
Jev Models & Reimplementations#community#jevbest
Community
firelex/jeffPython
firelex/jeffPython2kGitHub stars
DeepOpen is a non-autoregressive decision engine built on Laya that classifies into developer-defined types in one forward pass, with a language router picking between English, multilingual, and typed-decision checkpoints.
Jev Models & Reimplementations#community#jevbest
Community
deepopen-com/deepopenPython
deepopen-com/deepopenPython715GitHub stars
Decision models that answer several choice, score, and noul questions over a state with probabilities in one forward pass, with training, two inference backends, calibration tooling, and the benchmark suite in the repo.
Jev Models & Reimplementations#community#jevbest
Community
InternLM/Intern-DecisionPython
InternLM/Intern-DecisionPython402GitHub stars
Decision models that return typed probabilities for candidate labels from a state, instructions, and options, shipped as LoRA adapters plus a trained decision head over pinned Qwen base weights.
Jev Models & Reimplementations#community#jevbest
Community
Zefan-Cai/Open-JevPython
Zefan-Cai/Open-JevPython363GitHub stars
Open-weight 2B, 4B and 9B decision models that extend Jev's typed question format to photos, so an image plus a record can be checked against options you supply, with a probability on each and an explicit can't-tell. Runs on your own hardware.
Jev Models & Reimplementations#community#jevbest
Community
mohit67890/imajevPython
mohit67890/imajevPython253GitHub stars
A clone of browser-use's jev-ultrafast that swaps the hosted Jev decision call for Laya running locally through MLX, keeping the browser agent, action space and executor but making no paid decision API calls. Apple Silicon only.
Jev Models & Reimplementations#community#jevbest
Community
ipenywis/laya-ultrafastPython
ipenywis/laya-ultrafastPythonFrom the Jev guides
Original write-ups on what this entry is for.
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
One model in September, a dozen by October. Clef, Liquid d1, Perplexity, and OpenAI all ship the same typed-answer shape now. How to read the rival benchmarks and pick on your own task.
Jev Models & Reimplementations#decision-model#benchmarks#pricing
Read article
Hundreds of repositories carry the system-one tag, but the ecosystem is a dozen ideas in different sizes. A tour of the families, the licenses, and what a download does not include.
Jev Models & Reimplementations#open-models#hugging-face#local
Read article
From the Jev community
Posts from builders shipping with Jev right now.