Skip to content

New supported models

The proposal introduces 3 new models:

  • Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 - first big model on Gonka chain.
  • Qwen/Qwen3-32B-FP8 - fresh medium size model, replacement for Qwen/QwQ-32B
  • RedHatAI/Qwen2.5-7B-Instruct-quantized.w8a16 - replacement for Qwen/Qwen2.5-7B-Instruct (minor improvement to get rid of fully dynamic quantization)

New values estimations

The methodology for computing the thresholds is detailed in the thresholds_sep2025.ipynb. A reproducible version of this notebook is available in the project repository at /mlnode/packages/benchmarks/notebooks/thresholds_sep2025.ipynb.

The data used for our experiments is available for verification link. For maximum confidence, participants are encouraged to recompute the thresholds independently using the provided notebook.

  • All experiments were conducted using MLNode v3.0.9, which is the current version in the main branch..
  • Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 is proposed with --max-model-len 240000. This value is optimized for deployment on systems with 320GB of VRAM, allowing, for example, two instances of the model to run on a standard 8xH100 server.

Release process

If this proposal is approved, node operators will be able to modify their MLNodes config to switch to new models. Transition can be done asyncronously.

Detailed instructions for a seamless transition will be published on the official project website and announced through all relevant community channels.