MLX-VLM v0.7.4 Public — Adds support for DeepSeek V4.1 and multiple image generation enhancements

By the editors ／ Model: gemma-4-26b-a4b-it-qat-6bitAs of: 2026-09-29

MLX-VLM v0.7.4 has been released, bringing support for several new models and significant improvements to image generation workflows.

New Features and Improvements

Added support for DeepSeek V4.1 (vision and language models) and InternVL3.5.
Introduced the ability to generate multiple images per prompt via the num_images parameter.
Implemented batched image generation for qwen_image, flux2, and ideogram4.
Added support for Qwen3.5 MoE text and LFM2 encoder/ColBERT.
Added support for gpt-oss with reasoning and content parsing for harmony channels.
Improved image generation by using model-based sampling defaults.

Bug Fixes

Fixed audio request failures for MiMo-V2.6 in server multi-threaded environments.
Fixed Qwen3.5-arch left padding issues for batches and Qwen3-VL position embeddings with quantized pos_embed.
Resolved tool-parser failures involving boolean property schemas, untyped Qwen3-Coder objects, and unclosed parameters.
Fixed issues with sampling for small top-p and very low temperature values.
Fixed a bug in the Gemma 4 audio encoder regarding past frame attention.
Resolved issues with interleaved image order in shared prompt formatting and KV cache packing for certain quantized models.

Sources