NitrAI

/

OpenGCM-v2

like 0

Subject: Response regarding the flags raised by "solhost" on NitrAI repositories
​Dear Hugging Face Moderation Team,
​I am writing to formally address and clarify the allegations made by the user "solhost" regarding the NitrAI account and our model releases, specifically OpenGCM-v2.
​We strongly reject the claims of benchmark fabrication, bot activity, or bad-faith naming, and would like to provide clear context for each point raised:
​1. Regarding the Naming ("OpenGCM") and Brand Ownership
The acronym "GCM" stands for a generic technical description that we defined early on. The prefix "Open" is widely used across the entire open-source AI ecosystem (e.g., OpenELo, OpenCoder, OpenR1) to signify open-weights versions of models.
The user points to an Ollama repository page (ollama.com/OpenGCM) created very recently. However, we have been working on our architecture independently. To completely eliminate any potential user confusion and to show goodwill toward open-source collaboration, we are perfectly open to renaming our future weights/repositories if the moderation team deems it necessary, though we maintain that generic open-source naming conventions should remain open to the community.
​2. Regarding the Deletion of "v1" and Project Evolution
The user asks why "OpenGCM-v1" does not exist publicly anymore. The answer is standard development practice: v1 was an early experimental checkpoint that suffered from gradient instabilities. As an independent developer with limited infrastructure, maintaining obsolete, sub-optimal checkpoints is counterproductive. We deleted the deprecated v1 weights to save repository space and focused entirely on the heavily optimized v2 architecture.
​3. Regarding the "Absurd" 9B Benchmarks and Technical Validity
The claim that a 9B model cannot hit high scores or compete in specific areas with frontier models ignores recent breakthroughs in the open-source community (such as deep distillation, aggressive quantization filtering, and custom architectural tuning like Unsloth-optimized kernels). Our benchmarks were recorded under standard evaluation frameworks. We welcome any community member to pull the model weights, run independent evaluations via standard pipelines (like Lighteval or lm-evaluation-harness), and verify the replication of the output logic.
​4. Regarding the "Bot Account" Accusation
NitrAI is an active, human-operated project driven by genuine research and iteration. The sudden surge in trending metrics reflects the community's organic interest in highly efficient, smaller reasoning models capable of running on consumer hardware. We do not engage in any form of metric manipulation or "syphoning."
​Conclusion:
We respect the OpenGCM project team and their work, but open-source codebases frequently share similar acronyms and open-source methodologies. We are ready to cooperate with Hugging Face staff to adjust any metadata or naming to maintain platform integrity, but we request that these unverified, competitive-driven flags be dismissed.
​Sincerely,
NitrAI Team

To NitrAI,

Thanks for clearing up the situation.

At first we thought you were some bot account, but you have proven us wrong (for the most part).

I would still like to raise a flag:

• I still think some of these benchmarks are semi-unlikely, especially after reviewing the model on my machine.

After clearing up the bot allegation:

It seems much more likely the naming was just a coincidence, as bots may scrape websites (such as Ollama) for names.

We respect the NitrAI team and their contributions to the AI community.

Before closing this report, I would still like to wait for the head of OpenGCM to have a say on this situation.

Sincerely,
solhost, Co-Founder of the OpenGCM project.
Feel free to contact us via:

• Email: 2pst5lgbr@mozmail.com
  And our organizations:
  https://ollama.com/OpenGCM
  https://huggingface.co/OpenGCM
  Thanks.

Hi solhost,
​Thank you for the constructive update and for clearing up the misconceptions regarding our team and the naming coincidence. We appreciate your respect and fully return it - open-source collaboration thrives when teams communicate directly rather than assuming the worst.
​Regarding the benchmarks you tested, it is worth noting that the stock Qwen3.5-9B architecture is already exceptionally powerful out of the box, comfortably performing on par with frontier lightweight models like Gemini 3.1 Flash-Lite in many standard logic and coding tasks. When you combine such a strong base architecture with high-quality, dense fine-tuning traces, the model naturally pushes the upper limits of what a 9B parameter model can achieve. This is why the metrics might seem surprisingly high at first glance compared to older generation models.
​Additionally, please keep in mind that local performance can vary significantly depending on the exact inference parameters (such as system prompt alignment, temperature, top_p, and the specific quantization format used). Since our model was heavily optimized via QLoRA with specific target modules, slight deviations in the local setup or using a different prompt template can sometimes impact the reasoning chain or the final benchmark score. We stand by our evaluated metrics under standard evaluation harnesses, but we always welcome independent community deep-dives and feedback.
Best regards,
NitrAI Team

Dear NitrAI,

Once again thanks for taking the time out to reply. I appreciate the explanation that you have given in relation to your team and your development process.

On reflection of the matter at hand, my issue with the situation is not with the account but the project name itself. Our OpenGCM organization along with its first official release came prior to your OpenGCM-v2 release, and we intend on moving forward with our OpenGCM project in the coming years even with the upcoming Mark releases. Hence, the reason I am worried that having two different projects under the same name will lead to confusion on the user front.
Earlier, you said that you would be open to changing the names of your future repositories in case there arises any confusion for the sake of clarity. I just wanted to know if you would be willing to discuss this point with us and not leave it to platform moderators to decide.
Not trying to claim any malintent on your part; just trying to avoid confusion between the two projects.

Looking forward to hearing from you.

Hi TeamNull,
​Thank you for your response. While we appreciate the professional tone, we have to clarify the chronological facts regarding the project timelines.
​Our project, architecture, and the OpenGCM naming structure were already fully established, documented, and active on June 21, 2026. At that time, there was no public footprint, model release, or active repository under this name from your organization. As verified by the official Registry API and website metadata, your repository upload occurred later (around a week ago), well after our original launch.
​Furthermore, inspection of the manifest and blob metadata configuration confirms that your gcmmark1 model uses the standard public Qwen2/Qwen2.5 architecture and Q4_K_M quantization format rather than a long-standing custom historical architecture.
​Therefore, our team was the first to conceptualize, brand, and publicly launch a project under this specific name. While we sincerely respect your future roadmap and wish your team the best with your upcoming releases, we cannot alter the naming of our existing or directly related future repositories, as it represents our own established development history.
​To avoid any user confusion on the platform moving forward, we highly recommend adding distinct organizational or branding tags to your upcoming releases to differentiate your work from ours.
​We are glad we could clear this up via open dialogue, and we now consider this matter fully resolved.
​Best regards,
NitrAI Team