MF-1

MF-1 (Multimodal Flow) is an open-weights, 1.6-billion-parameter foundation model developed by Horizon Robotics in collaboration with Huazhong University of Science and Technology and Beijing Jiaotong University. Released in October 2026, the model departs fundamentally from the standard design pattern of modern multimodal systems. Rather than bolting a discrete vision transformer to an autoregressive text model, MF-1 establishes a unified representation for both language and images using continuous Flow Matching.

The model is built to handle language modeling, visual question answering, text-to-image generation, and image editing natively within a single set of weights. Trained on 150 billion pre-training tokens, MF-1 punches far above its parameter class, scoring 82.8 on GenEval and 75.3 on DPG-Bench. For developers seeking a capable local AI model with 1.6B parameters in 2026, MF-1 represents an efficient, all-in-one vision-language engine capable of both multimodal comprehension and generation.

Because it eliminates separate vision encoders and discrete vector-quantization bottlenecks, MF-1 presents unique performance characteristics for on-device deployment. Its compact footprint makes it an ideal candidate for edge devices, consumer workstations, and embedded robotics hardware where memory bandwidth and capacity are constrained.

MF-1 uses a dense 1.6-billion-parameter architecture built on a chunk-causal flow backbone. Most vision-language models rely on a dual-component pipeline: a visual encoder (such as CLIP or SigLIP) compresses pixels into embeddings, which are then projected into a causal transformer backbone alongside text tokens. Systems that generate images typically reverse this via discrete visual tokens (VQ-VAE/VQ-GAN) or pipe the output into a separate diffusion model.

MF-1 replaces this disjointed setup with a continuous generative flow:

• Continuous Hyperchunks: Both text sequences and image rasters are organized into ordered continuous blocks termed "hyperchunks." This format maintains natural language order for text while preserving 2D spatial relationships for images, removing the need for a discrete visual quantization bottleneck.
• Flow Matching Objective: The backbone learns a unified vector field across these hyperchunks using Flow Matching. Instead of predicting the next discrete token in an image codebook, it generates continuous visual and textual representations through deterministic vector trajectories.
• Joint Attention with Modality-Specific FFNs: The architecture applies joint self-attention across cross-modal inputs to allow dynamic interaction between text and image signals. Modality-specific feed-forward networks (FFNs) then process the specialized features of text and visual data independently.
• Inference Mechanics: During training, the architecture optimizes parallel predictions across multiple target chunks. At inference time, the model generates hyperchunks sequentially, balancing the autoregressive nature of continuous text streaming with the multi-step trajectory integration required for image generation.

Because the model is dense, every parameter is engaged during a forward pass. This gives MF-1 consistent computational characteristics without the routing overhead or uneven memory spikes associated with sparse Mixture-of-Experts (MoE) architectures.

MF-1 handles both perception and generation tasks within one weight distribution. The training recipe enables several key capabilities:

• Interleaved Text and Image Generation: MF-1 can output text responses, generate novel images from scratch, or produce interleaved multimodal documents without switching pipelines.
• Precision Image Editing: By modeling visual data as continuous vector fields, MF-1 natively performs instruction-based visual editing, localized inpainting, and stylistic transformations while adhering to natural language prompts.
• Complex Compositional Generation: The model scores 82.8 on GenEval and 75.3 on DPG-Bench. These scores demonstrate high fidelity in spatial positioning, attribute binding, and multi-object generation, matching or outperforming several dedicated diffusion models twice its size.
• Visual Understanding and Reasoning: It acts as an on-device visual assistant, extracting information from screenshots, analyzing diagrams, and answering grounded questions about visual scenes.
• Local Robotics and Edge AI: Given Horizon Robotics' core focus on edge computing and robotics, MF-1 serves as a lightweight brain for embedded platforms that require zero-shot perception paired with visual predictive planning.

Due to its 1.6B dense footprint, the MF-1 hardware requirements are remarkably modest, making it accessible on almost any modern consumer GPU, integrated SoC, or older enterprise accelerator.

The precise memory footprint depends on the execution mode (pure language generation vs. multi-step flow integration for images) and the quantization format:

• FP16 (Unquantized): ~3.5 GB VRAM for model weights; 5.5 to 6.5 GB VRAM recommended to hold working KV caches and flow matching trajectory buffers.
• Q8_0 (8-bit Quantization): ~1.8 GB VRAM for model weights; 3.5 to 4 GB VRAM recommended.
• Q4_K_M (4-bit Quantization): ~1.1 GB VRAM for model weights; 2.5 to 3 GB VRAM recommended.

To run MF-1 locally, practically any machine with a dedicated GPU or unified memory will suffice:

• Entry-Level Consumer GPUs: An NVIDIA GeForce RTX 3060 (12GB), RTX 4060 (8GB), or AMD Radeon RX 7600 XT runs MF-1 comfortably at full FP16 precision with plenty of headroom for surrounding application memory.
• High-End Desktop Workstations: On an RTX 4090 or RTX 4080, MF-1 operates entirely in ultra-fast L2/VRAM cache hierarchies, yielding extreme throughput.
• Apple Silicon: Macs equipped with M2, M3, or M4 chips (including base 8GB or 16GB configurations) run quantized builds of MF-1 effortlessly via unified memory architectures.

Choosing the best quantization for MF-1 depends on whether your workload is text-centric or vision-centric:

• Q4_K_M: The optimal balance for text inference, document summarization, and basic visual Q&A. It cuts the memory footprint down to roughly 1.1 GB with negligible degradation in linguistic perplexity.
• Q8_0 or FP16: Recommended if you rely heavily on the continuous Flow Matching pipeline for image generation and fine-grained visual editing. Aggressive 4-bit quantization can occasionally introduce variance into the numerical solvers used during flow trajectory integration, leading to subtle visual artifacts or lower prompt fidelity in generated images.

When deployed on an NVIDIA RTX 4070 (12GB), expect the following MF-1 performance figures:

• Text-Only Generation: MF-1 tokens per second regularly surpass 110 to 140 tok/s at 4-bit precision, and 75 to 90 tok/s at FP16.
• Continuous Image Synthesis: Generating a standard visual hyperchunk via the Flow Matching sampling loop typically completes in 1.2 to 2.8 seconds, depending on the integration steps assigned to the continuous vector solver.

If you are looking at how to run 1.6B model on consumer GPU setups via standard frameworks, containerized runtimes and execution backends supporting flow matching layers provide the cleanest path. For standard textual inference, Ollama and llama.cpp offer straightforward execution once model weights are converted to GGUF format, though full multimodal image synthesis requires runtimes that implement the model's unified flow matching ODE/SDE solvers.

Evaluating MF-1 against alternative compact architectures highlights the trade-offs of its unified continuous flow design:

Qwen2.5-VL-1.5B is a standard-bearer in the sub-2B vision-language tier. It pairs a discrete ViT encoder with an autoregressive language backbone.

• Perception: Qwen2.5-VL excels at raw document parsing, OCR, and granular visual reasoning due to its high-resolution dynamic patch processing.
• Generation: Qwen2.5-VL cannot generate or edit images; it is strictly an input-vision, output-text model.
• Takeaway: Choose Qwen2.5-VL-1.5B if your workload is purely OCR, document understanding, and text extraction. Choose MF-1 if you need bidirectional capability, specifically generating or editing visual content alongside text in an identical memory footprint.

Meta's Chameleon explored early native multimodal tokenization by quantizing images into discrete tokens using a visual codebook.

• Visual Artifacts: Chameleon's reliance on discrete visual quantization created a severe bottleneck, resulting in lower generation quality and training instability.
• Efficiency: MF-1 circumvents this by using continuous Flow Matching over continuous hyperchunks. As a result, MF-1 (at 1.6B) scores 82.8 on GenEval, outperforming larger discrete-token multimodal architectures while requiring a fraction of the compute and VRAM.
• Takeaway: MF-1 proves that continuous representations solve the visual quality loss inherent in discrete codebook architectures, providing superior generation quality at a significantly lower parameter count.