Add DeepSeek V4 Pro and Qwen3.8-27B Models
Author: Aung Myat Moe Created: 2026-08-19 06:34 UTC Closes: 2026-08-27 06:34 UTC Language: EN Votes: 2 Avg. Bid: 0.00 GNK
Expedited Update Params (Inference), Register Model
Full Proposal
Motivation The Gonka network relies on offering top-tier, high-demand AI models to remain competitive and attract a steady stream of inference requests. The recent advancements in open-weight models have brought two standout performers to the forefront:
DeepSeek V4 Pro: DeepSeek's latest iteration offers state-of-the-art reasoning, coding, and mathematical capabilities, often rivaling or exceeding proprietary counterparts at a fraction of the parameter cost. There is significant developer demand for DeepSeek models, as evidenced by the success of our recent addition of DeepSeek V4 Flash 0731.
Qwen3.8-27B: The Qwen series continues to excel in multilingual proficiency, robust context handling, and high-efficiency inference. The 27B parameter size is the "sweet spot" for decentralized inference, offering excellent capabilities while remaining hardware-friendly enough to ensure a broad distribution of capable nodes within our network.
Adding these two models will directly increase the network's utility, expand our user base, and drive higher computational volume through the network.
2. High-Level Solution
We propose registering both DeepSeek V4 Pro and Qwen3.8-27B to the Gonka model registry. This involves updating the network parameters to formally accept inference workloads for these models.
Register Model: Formally add the model identifiers and configurations to the chain.
Update Params (Inference): Adjust the inference parameters to allocate appropriate weight_scale_factor multipliers for these models based on their computational requirements.
3. Implementation Details
If this proposal passes, the following on-chain actions will be executed:
Register deepseek-v4-pro: Add the official DeepSeek model architecture configurations.
Register qwen3.8-27b: Add the official Qwen architecture configurations.
Set Weight Scale Factors:
Set the weight_scale_factor for deepseek-v4-pro to reflect its heavy computational demands relative to smaller models.
Set the weight_scale_factor for qwen3.8-27b to appropriately reward nodes processing this medium-to-large model. (Exact parameter multipliers to be finalized during community discussion prior to the final on-chain vote).
4. Open Questions for the Community Call
What should the exact weight_scale_factor be for these models to ensure fair compensation for node operators without pricing out end-users?
Should we gradually phase out any older, underutilized models (similar to the removal of older Kimi models) to free up bandwidth and storage for node operators?
Comments (7)
💬 Slava MyGonka
2026-08-19 08:25 · 👍 1 · 👎 0
Активное участие в добавлении новых моделей принимают эти ребята: https://registry.kaitaku.ai/
https://t.me/baridoka
Думаю, тебе было бы интересно с ними пообщаться.
Именно они разрабатывали коэффициент для добавления Deep Seek.
💬 Aung Myat Moe
2026-08-19 08:30 · 👍 1 · 👎 0
Yup I send message to them and I checked their images and most of them are optimized properly and new ds is just putting the entry
💬 Slava MyGonka
2026-08-19 08:34 · 👍 1 · 👎 0
Я не могу найти их расчеты. Они где-то есть на GitHub. Перед добавлением подели они тестировали эту модель на разном оборудовании. Результаты тестов есть в отчете на GitHub.
Коэффициенты при добавлении модели очень важны.
💬 Slava MyGonka
2026-08-19 08:38 · 👍 1 · 👎 0
https://github.com/kaitakuai/experiments вот, ребята подсказали из Комьюнити. Ты это читал?
💬 Slava MyGonka
2026-08-19 07:14 · 👍 0 · 👎 0
Я считаю, что сеть пока не готова к приему моделей, которые не помещаются на 8*Н100. Ты видел, что было с GLM? То же самое ждет и DeepSeek V4 Pro, т.к. она слишком большая для доступных GPU.
А модель Qwen3.8-27B помещается на 1×H100 80 GB
Мы уже это проходили. Слишком маленькие модели позволяют заниматься читингом. Поэтому от них отказались.
💬 Aung Myat Moe
2026-08-19 08:05 · 👍 0 · 👎 0
Nvx version can accept and it’s acceptable at performance and accuracy bro.
And the small model is dense qwen is at opus level and competing with Fabel .
To fix cheating governance need to update like TAO and make sure to put devshard as validate
I am planning to make llm as verifier with new proposal before the research is done it should be able to do it
💬 Aung Myat Moe
2026-08-19 23:11 · 👍 0 · 👎 0
Ok