Add EximiusLabs/fusion-embedding-2-2b-preview (multimodal embedding: text, image, video, audio)

[abtonmoy]

Adds a recipe for EximiusLabs/fusion-embedding-2-2b-preview, a 2.8B multimodal embedding model that places text, images, video and audio in one 2048-d space (frozen Qwen3-VL-Embedding-2B base + frozen Qwen2.5-Omni audio tower + token-gated adapters).

The model serves through an out-of-tree plugin registered via the vllm.general_plugins entry point; it ships inside the fusion-embedding PyPI package, so the install is uv pip install 'vllm[audio]==0.26.0' fusion-embedding and the launch is a plain vllm serve EximiusLabs/fusion-embedding-2-2b-preview --runner pooling (no trust-remote-code, no extra flags). Smoke-tested against vllm==0.26.0.

Notes for review:

• node scripts/build-recipes-api.mjs passes and renders public/EximiusLabs/fusion-embedding-2-2b-preview.json.
• Every number in the recipe is measured: parameter counts summed from safetensors metadata, context length from the base config, parity figures from executed comparison runs against the reference implementation (fp32 text cosine 0.999998+, image 0.999986+, audio 0.999999+; the served HTTP endpoint verified end to end).
• Single-GPU, single_node_tp with tp: 1, eager execution forced by the plugin (documented in the guide).

Happy to adjust tags, difficulty, or guide structure to fit conventions.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview	Aug 9, 2026 8:46am

[gemini-code-assist]

Code Review

This pull request introduces a configuration file and guide for the fusion-embedding-2-2b-preview multimodal embedding model. The feedback recommends replacing uv pip install with standard pip install for broader compatibility, increasing the minimum VRAM requirement to 8 GB to match standard hardware tiers, and wrapping the offline API input in a list to prevent potential type validation errors in vLLM.

[gemini-code-assist]

Using pip install instead of uv pip install as the primary command is recommended for general compatibility, as uv is an external tool that is not pre-installed in most Python environments.

      command: "pip install 'vllm[audio]==0.26.0' fusion-embedding"

[gemini-code-assist]

Increasing vram_minimum_gb to 8 aligns with standard hardware tiers (such as 8 GB GPUs) and the guide's recommendation of 'roughly 8 GB VRAM or more'. A 7 GB limit is not a standard GPU size and is extremely tight for loading ~5.7 GB of weights plus the CUDA context and vLLM overhead.

    vram_minimum_gb: 8

[gemini-code-assist]

Using standard pip install in the guide's installation instructions provides better out-of-the-box compatibility for users who do not have uv installed.

  pip install 'vllm[audio]==0.26.0' fusion-embedding

[gemini-code-assist]

Wrapping the multi-modal input dictionary in a list is the standard and most robust way to pass inputs to llm.embed in vLLM, preventing potential type validation errors in certain vLLM versions.

  out = llm.embed([{"prompt": "<|vision_pad|><|im_end|>",
                    "multi_modal_data": {"audio": (wav, sr)}}])

[abtonmoy]

Addressed the review: bumped vram_minimum_gb to 8 (matches the guide's prerequisites line and real hardware tiers) and wrapped the offline llm.embed input in a list, both in 3e50af9. Kept uv pip install for the install commands since that is the dominant convention across the existing recipes (285 occurrences vs 11 plain pip install); happy to switch if maintainers prefer otherwise. Validator still passes.