Pull down to go back
Gemma 4's Hidden Multi-Token Prediction Feature: Community Reverse Engineering Effort Underway
opensourcereddit

Gemma 4's Hidden Multi-Token Prediction Feature: Community Reverse Engineering Effort Underway

Gemma 4 隱藏的多令牌預測功能曝光：社群逆向工程進行中

A developer has successfully extracted Gemma 4's model weights and discovered the AI model has a multi-token prediction (MTP) capability that wasn't publicly documented. Now they're calling on the community—especially C++ experts—to help reverse engineer the MTP feature from compiled TFLite files back into a usable PyTorch format. The extracted files are already available on HuggingFace, and this could unlock new possibilities for how the model generates text. If you've got reverse engineering skills, this is your moment to contribute to something that could benefit the entire open-source AI community.

Read original →

Keywords

reverse engineeringmodel weightsMTPTFLiteneural networkcommunity contribution

Related Articles

TRELLIS.2 image-to-3D now runs on Mac (Apple Silicon) - no NVIDIA GPU needed
Open SourceRead

TRELLIS.2 image-to-3D now runs on Mac (Apple Silicon) - no NVIDIA GPU needed

I ported Microsoft's TRELLIS.2 to run on Apple Silicon via PyTorch MPS. The original depends on five CUDA-only compiled extensions (flex_gemm, flash_attn, o_voxel, cumesh, nvdiffrast) that have no Mac equivalent. Wrote replacement backends from scratch: Pure-PyTorch sparse 3D convolution (replacing flex_gemm), Python mesh extraction using spatial hashing (replacing CUDA hashmap ops in o_voxel), SDPA attention for sparse transformers (replacing flash_attn), and GPU-accelerated trilinear interpolation (replacing cumesh and nvdiffrast).

Reddit
Show HN: Run TRELLIS.2 Image-to-3D generation natively on Apple Silicon
Open SourceRead

Show HN: Run TRELLIS.2 Image-to-3D generation natively on Apple Silicon

I ported Microsoft's TRELLIS.2 (4B parameter image-to-3D model) to run on Apple Silicon via PyTorch MPS. The original requires CUDA with flash_attn, nvdiffrast, and custom sparse convolution kernels: none of which work on Mac. I replaced the CUDA-specific ops with pure-PyTorch alternatives: a gather-scatter sparse 3D convolution, SDPA attention for sparse transformers, and a Python-based mesh extraction replacing CUDA hashmap operations. Total changes are a few hundred lines across 9 files.

Hacker News
157
Which Qwen 3.6 Quantization Works Best for M5 Pro with 24GB RAM?
Open SourceRead

Which Qwen 3.6 Quantization Works Best for M5 Pro with 24GB RAM?

A user with an M5 Pro machine featuring 24GB of RAM is trying to figure out the best quantization level for running Qwen 3.6 via Ollama. They're unsure whether to go with Q4 (4-bit quantization) but can't find a solid Q3 (3-bit) alternative. They're looking for recommendations on which version would run smoothly on their setup without compromising too much on quality.

Reddit