strands-qwen3-vl-2b

A Strands-Agents-expert Qwen3-VL-2B that keeps learning after deployment.

Qwen/Qwen3-VL-2B-Instruct post-tuned (LoRA r=32, merged) on the full public codebase and docs of strands-agents and strands-labs (12 repos, ~5.6M tokens), designed to be wrapped with the strands-slm plastic self-learning layer: surprise-gated weight updates at inference, EMA-bounded plasticity, and a provable off-switch (reset() is bit-identical to this base).

Code, demo notebook, and full experimental record: github.com/cagataycali/slm

Use — self-learning (recommended)

pip install strands-slm

As a Strands Agents model provider — every agent turn can change the weights:

from strands import Agent
from strands_tools import shell
from slm import SLM

model = SLM("cagataydev/strands-qwen3-vl-2b")
agent = Agent(tools=[shell], model=model)
agent("use the shell tool to run: echo hello")   # this turn updated the weights

Or drive the learning loop directly:

from slm import StrandsPlasticQwen

m = StrandsPlasticQwen.from_pretrained("cagataydev/strands-qwen3-vl-2b", device="cuda")
print(m.chat("How do I create a custom tool in Strands Agents?"))

for doc in your_stream:
    m.observe(doc, learn=True)   # predicts; if surprised, rewrites its fast weights
m.reset()                        # bit-exact back to this base

Use — plain transformers (frozen expert)

from transformers import AutoModelForImageTextToText, AutoProcessor
model = AutoModelForImageTextToText.from_pretrained("cagataydev/strands-qwen3-vl-2b", dtype="bfloat16", device_map="cuda")
proc = AutoProcessor.from_pretrained("cagataydev/strands-qwen3-vl-2b")

Results

Measured on a single L40S, seed-replicated. The base is never updated by the plastic layer.

claim	evidence
Domain expert	strands probe NLL 4.85 → 2.22, 8/8 probes improved; corpus holdout NLL 1.848 → 1.239
Learns while running	continual OOD stream NLL 6.18 → 5.37, pure inference
Does not forget	strands expertise after OOD learning: Δ −0.01
Agent competence grows	held-out tasks 0/4 → 4/4 after 18 curated lessons, 5/5 seeds
Fact memory	15/15 facts at 100% verbatim recall
Provable off-switch	reset() is bit-identical to this base, Δlogits = 0
Cost	+0.11–0.25 s/turn learning overhead

Full results, honest limitations, and the demo notebook: github.com/cagataycali/slm

License: Apache-2.0 (inherits Qwen3-VL license terms).