rskill-qwen35-4b-nf4

OpenRAL rSkill — Qwen3.5-4B natively-multimodal video-language model packaged as an NF4 bitsandbytes vlm rSkill (ADR-0047). Accepts RGB image or video frames plus a natural-language query; returns a text answer. No actuators. Apache-2.0.

Quick Start

ral skill install hf://OpenRAL/rskill-qwen35-4b-nf4

from openral_core.schemas import RSkillManifest

manifest = RSkillManifest.from_yaml("rskills/qwen35-4b-nf4/rskill.yaml")
assert manifest.kind == "vlm"
assert manifest.role == "s2"
assert manifest.quantization.extra["scheme"] == "nf4"
assert manifest.is_commercial_use_allowed is True

What It Does

Qwen3.5-4B is a natively-multimodal foundation model trained from scratch on interleaved text, image, and video tokens. Given an RGB image or video clip and a natural-language question, it returns a free-form text answer grounded in the visual content.

This rSkill declares kind: vlm and role: s2 because it is a pure perception component operating at S2 (slow-reasoning) rate (~0.2–1 Hz), not an S1 fast policy. It consumes camera frames and natural-language queries, emits text answers, and never drives ros2_control joints.

Representative queries for robot scene understanding:

• "What objects are on the table?"
• "Is the gripper clear of obstacles?"
• "Describe the relative positions of the cup and bowl."
• "Has the pick-and-place task completed?"

Why Qwen3.5-4B over Qwen2.5-VL-7B

	Qwen3.5-4B (this skill)	Qwen2.5-VL-7B
Parameters	4B	7B
VideoMME (w/ subs.)	83.5%	~72%
MLVU	82.8%	~73%
VRAM at NF4	~2.5 GB	~3.3 GB
VRAM at BF16	~8 GB	~13 GB
Architecture	Hybrid linear-attn (3:1)	Full quadratic ViT+LLM
License	Apache-2.0	Apache-2.0

Qwen3.5-4B beats Qwen2.5-VL-7B on every video benchmark despite being 3B smaller. The 3:1 Gated DeltaNet / full-attention hybrid processes long video sequences far more efficiently — important for continuous robot camera streams. At NF4 it fits well within 8 GB VRAM alongside the S1 skill stack.

Architecture

Qwen3.5 uses a 3:1 hybrid attention stack: three Gated DeltaNet (linear-attention, O(n)) layers for every one full-attention layer. This reduces cost on long sequences significantly. The vision encoder is shared with Qwen3-VL. Key features:

• Native video support — temporal patch embedding, second-level event localization, up to 256K context (extensible to 1M)
• Spatial grounding — RefCOCO avg ~80.6; strong for "where is X?" queries
• 201-language support

Runtime

This rSkill ships a pre-quantized NF4 checkpoint as weights_uri (hf://OpenRAL/rskill-qwen35-4b-nf4): model.safetensors with an embedded bitsandbytes quantization_config (nf4, double-quant, bf16 compute). The sidecar loads it directly as 4-bit (~3.3 GB resident, no bf16 load spike), so it fits an 8 GB GPU with no loader workaround. source_repo records the SHA-pinned upstream Apache-2.0 model it was quantized from (provenance, §8).

Reproduce the checkpoint with tools/build_qwen_vlm_nf4_checkpoint.py (run in the sidecar venv):

$OPENRAL_QWEN_VLM_SIDECAR_VENV/bin/python tools/build_qwen_vlm_nf4_checkpoint.py \
  --source Qwen/Qwen3.5-4B \
  --out ~/.cache/openral/qwen35-4b-nf4-ckpt

It loads the upstream model once (forcing serial materialization so the bf16 pass fits 8 GB), saves the NF4 weights + processor, then verifies the checkpoint reloads directly as 4-bit and answers a smoke query.

The kind: vlm runtime is implemented (ADR-0047) as a read-only reasoner tool, not an ExecuteSkill (a scene VLM produces text, not actions):

• Sidecar: tools/qwen_vlm_sidecar.py boots the NF4 model in its own venv and serves a ZMQ REQ/REP + msgpack protocol. Provision it separately and point at it with OPENRAL_QWEN_VLM_SIDECAR_VENV (or let the backend auto-spawn it on first query).
• Backend: openral_runner.backends.gstreamer.qwen_scene_vlm.QwenSceneVlm is the node-side ZMQ client; build_scene_vlm(manifest) builds it from this manifest. The node-side client deps (pyzmq + msgpack) install with uv sync --group qwen-vlm.
• Service node: openral_perception_ros.scene_vlm_node subscribes the cameras and serves /openral/perception/query_scene (openral_msgs/srv/QueryScene).
• Reasoner tool: the LLM sees the read-only query_scene tool when the reasoner is launched with scene_query_available:=true. It asks open-ended scene-state questions ("has the robot grasped the mug?", "is the task complete?") and the answer feeds the next reasoning tick.

Validated live

The sidecar + backend + query_scene path was run end-to-end on an NVIDIA RTX 4070 Laptop (8 GB): NF4 Qwen3.5-4B loads to ~3.3 GB resident, and real image queries return correct answers — including the task-verification use case ("Has a robot gripper grasped any object?" → "No", grounded in the frame). Covered by the GPU-gated tests/unit/test_qwen_scene_vlm.py::test_e2e_query_coco_sample (set OPENRAL_QWEN_VLM_SIDECAR_VENV), not asserted blind.

8 GB load note. Deploying the pre-quantized weights_uri loads the 4-bit weights directly (~3.3 GB, ~6 s) with no workaround — the clean 8 GB path. The workaround only matters when quantizing at load from the raw upstream (the build step, or --model Qwen/Qwen3.5-4B): transformers 5.x's parallel loader materializes weights in bf16 on-GPU before bitsandbytes quantizes, and the 4-way-concurrent ~7.4 GB transient OOMs an 8 GB card, so the sidecar forces serial materialization (core_model_loading.GLOBAL_WORKERS = 1) + PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True. The sidecar auto-detects which path applies. The Gated-DeltaNet fast kernels (fla / causal-conv1d) are optional — without them transformers uses a slower torch fallback (the model still loads and answers). The model loads via AutoModelForImageTextToText (it registers as Qwen3_5ForConditionalGeneration).

Benchmark Numbers

Benchmarks below are paper-reported (Qwen team, February 2026); reproduced_locally: false in the eval JSON.

Benchmark	Qwen3.5-4B	Qwen3.5-9B
VideoMME (w/ subtitles)	83.5%	84.5%
VideoMME (w/o subtitles)	76.9%	78.4%
VideoMMMU	74.1%	78.9%
MLVU	82.8%	84.4%
MVBench	71.2%	74.4%
LVBench	66.4%	70.0%
MMMU	77.6%	78.4%
RefCOCO avg	80.6%	81.3%
LingoQA (driving / spatial)	74.4%	80.4%

Supported robots and embodiments

This scene VLM is embodiment-agnostic — it reasons about camera frames and emits text, never actuator commands, so it imposes no kinematic requirement. The only hardware dependency is an RGB camera stream of at least 336×336. All in-tree OpenRAL embodiment tags are therefore listed in rskill.yaml (aloha, franka_panda, g1, google_robot, gr1, h1, mobile_base, openarm, panda_mobile, pusht, rizon4, sawyer, so100_follower, so101_follower, ur10e, ur5e, widowx) so any robot with a compatible camera can install it and expose the reasoner's query_scene tool. It pairs with any S1 VLA policy: the VLA acts, this VLM verifies (e.g. "did the grasp succeed?").

Sensors and Observation Contract

Direction	Key	Modality	Shape / format	Notes
in	any RGB camera	RGB image or video	min 336 × 336	vla_feature_key intentionally omitted
in	query	text	natural language	scene question, grounding query, or task-completion check
out	answer	text	free-form	grounded text response; adapter parses to SceneQueryResult

The model emits no action chunks and has no proprioception contract.

Manifest Summary

Field	Value
name	OpenRAL/rskill-qwen35-4b-nf4
version	0.1.0
license	apache-2.0
role / kind	s2 / vlm
runtime	pytorch
quantization.dtype	int4
quantization.extra.scheme	nf4
weights_uri	hf://OpenRAL/rskill-qwen35-4b-nf4 (pre-quantized NF4)
min_vram_gb.bf16	8.0 GB
min_vram_gb.int4	2.5 GB
latency_budget.per_chunk_ms	3000 ms
actions	query

License

The rSkill package metadata and README are OpenRAL project files under Apache-2.0. The wrapped Qwen3.5 weights are released by the Qwen Team under Apache-2.0, permitting commercial use. No OPENRAL_ALLOW_NONCOMMERCIAL=1 flag is needed.