TokenRhythm opens limited beta for NeoHorse-1.5-Flash Preview, a 180B agent model with 6B active
TokenRhythm released NeoHorse-1.5-Flash Preview on November 5, a 180B-parameter model with 6B active parameters aimed at agent task execution and software engineering. Access is limited to a beta via the company's WeChat account, and no weights, model card or license have been published.
TokenRhythm, a Chinese AI infrastructure company that runs a multi-model routing platform, released NeoHorse-1.5-Flash Preview on November 5. The model has 180 billion total parameters with 6 billion active, and is available only through a limited beta. Users apply via the company's official WeChat account.
Key specifications
- Developer: TokenRhythm
- Release date: November 5, 2026 (limited beta)
- Parameters: 180B total, 6B active
- Base model: trained on Qwen3.8-Flash, according to TokenRhythm
- Positioning: agent-native model for task execution and software engineering
- Context window: not yet disclosed
- Pricing: not yet disclosed
- Weights, model card, license: none published on TokenRhythm's GitHub or Hugging Face pages as of publication
Company-reported results
TokenRhythm compared the model against its Qwen3.8-Flash base. According to the company, overall scores rose by 9.23 points on Agents' Last Exam and by 5.92 points on Natural Language to Repository. It also says physical-science tasks improved by more than 20 points, with gains in developer tooling and testing, web development, and security and cryptography.
These are company-run evaluations. Independent results have not been published.
Recursive self-improvement, second round
TokenRhythm presents the release as the second round of its recursive self-improvement work. After NeoHorse-1 shipped in September, the company says it put that model back into its model pool, collected execution feedback from real tasks, and used it to train the next model.
The company describes the mechanism as follows: its routing system supplies signals about what capabilities tasks require, and running several frontier models on the same task produces comparable results. Gaps between NeoHorse's output and those results are turned into training signals. The team names three techniques behind the process: multi-model differential diagnosis, adversarial environment evolution and swarm online learning.
The loop also relies on OpenSquilla, TokenRhythm's open-source agent product. An update released the day after the model picks a model tier by task intent, keeps that tier when a task continues and steps up a tier on retries.
Not an open-weight release
TokenRhythm's public GitHub and Hugging Face pages confirm the earlier generation, not the new preview. NeoHorse-1 was released on September 7 as 4B and 9B models post-trained from Qwen3.5 under the Apache 2.0 license, with a technical report and quantized versions. A 4B decision model, NeoHorse-Jev, followed in late September.
As of publication, NeoHorse-1.5-Flash Preview had no downloadable weights, model card or license on those pages, so it should not be treated as an open-weight release.
The report draws on coverage from QbitAI and Jiemian, as well as TokenRhythm's GitHub repository and Hugging Face models page.
What this means
NeoHorse-1.5-Flash Preview is a larger model than the 4B and 9B NeoHorse-1 releases, with 6B active parameters out of 180B total. Its benchmark gains are measured against its own Qwen3.8-Flash base and were run by TokenRhythm, so they show the direction of the company's training method but not how the model compares with others. Independent evaluation will be needed to test the claims.
The release also shows how TokenRhythm connects its products: the routing platform supplies task signals, OpenSquilla supplies agent execution, and both feed training. Unlike NeoHorse-1, the preview comes with no open weights, license or model card, and access depends on an application through WeChat. Context window and pricing are not yet disclosed, so developers cannot yet assess it for production use.
Related Articles
Tencent releases Hy4: 770B total parameters, 49B active, with native vision and computer use
Tencent has released Hy4, with 770B total parameters and 49B active parameters. It is the first Hy model with native vision and computer-use support. Weights are due to be open-sourced soon, and a lower-cost Hy4 flash is coming.
Nous Research releases Arete-27B, a Qwen3.8-27B fine-tune that cuts propaganda rate from 22% to 0.9%
Nous Research fine-tuned Alibaba's Qwen3.8-27B on about 53 million tokens to remove Chinese state propaganda from its answers. According to Nous, the resulting Arete-27B model drops the propaganda rate from 22 percent to 0.9 percent while keeping general capabilities intact. Weights, test set and training data are public under Apache 2.0.
PrunaAI releases P-H3-Image: 8 MP image generation and editing in 8 steps, built on MiniMax-H3
PrunaAI has released P-H3-Image, an image model based on MiniMax-H3 that renders 8-megapixel images in 8 steps with no classifier-free guidance. It handles text-to-image and editing with up to 8 input images, and bf16 inference needs about 101 GB of GPU memory.
OpenCSG Agentic-27B: 27.36B-Parameter Tool-Use Model Runs on One A800, Commercial Use Needs Permission
OpenCSG has released Agentic-27B, a 27.36B-parameter model for multi-step tool use, built on the dense Qwen3.8-27B. The company reports it runs on a single 80GB NVIDIA A800, but commercial use requires authorization under the OpenCSG license, and all benchmark results are OpenCSG's own.
Comments
Loading...