Qwen 3.5 — Alibaba's Latest LLM Rivaling GPT-4o on Your Local Fleet

Qwen 3.5 by Alibaba — run Qwen 3.5 (the latest and most capable Qwen model) across your local device fleet. Qwen 3.5 rivals GPT-4o and Claude 3.5 on reasonin...

Security Scan

The skill's instructions are broadly coherent for running Qwen locally via an Ollama Herd router, but metadata and manifest inconsistencies plus an install-from-PyPI step mean you should verify the upstream package and repo before installing.

Purpose & Capability

The SKILL.md describes a local fleet router (herd/herd-node), local HTTP endpoints, and model pulls — all consistent with a 'run Qwen on your fleet' skill. However the registry metadata reported earlier (no required bins/config paths) conflicts with SKILL.md metadata that lists required binaries (curl/wget, optional python3/pip) and configPaths (~/.fleet-manager/...). The required binaries and config paths in SKILL.md are plausible for this purpose, but the mismatch between the published registry fields and the SKILL.md is an inconsistency worth noting.

Instruction Scope

SKILL.md only instructs installing an open-source PyPI package (ollama-herd), running local binaries (herd, herd-node), and calling localhost HTTP endpoints (port 11435). It does not instruct reading arbitrary system files or exfiltrating data to third parties. It does reference and warn about ~/.fleet-manager/, which indicates the skill will interact with that local directory; this is consistent with a fleet manager.

Install Mechanism

There is no platform-level install spec; installation is via the SKILL.md recommendation to 'pip install ollama-herd' (a PyPI package). Installing from PyPI is a common practice but carries the usual supply-chain risk: you should verify the PyPI package name, its maintainer, and the linked GitHub repository before installing. No arbitrary archive downloads or obscure URLs are present in SKILL.md.

Credentials

The skill does not request credentials or environment variables in the registry or in SKILL.md. The SKILL.md metadata asks for access to ~/.fleet-manager files (configPaths), which is proportional for a fleet/router tool but means the skill may read/write persistent local state under that directory.

Persistence & Privilege

The skill is not 'always' enabled and does not request platform-wide privileges, but SKILL.md metadata lists configPaths under ~/.fleet-manager. That implies persistent local data (latency.db, logs/herd.jsonl) will be used; the registry's top-level manifest did not declare those config paths, creating an inconsistency. Running herd/herd-node will open a local network service (port 11435) and create persistent state — be aware of network exposure and where state is stored.

What to consider before installing

This skill appears to be what it says (a local Ollama Herd router for Qwen models) but has a few red flags to check before installing: 1) Verify the PyPI package 'ollama-herd' and the GitHub repo (https://github.com/geeks-accelerator/ollama-herd) match and are maintained by a trusted source — review the package contents or source code. 2) Note the SKILL.md expects curl/wget and optionally python3/pip and will create/use ~/.fleet-manager/* files; inspect those files for sensitive data and consider changing permissions. 3) Running 'herd' exposes a local HTTP port (11435) — run behind a firewall or bind to localhost only if you want to avoid network exposure. 4) Prefer installing in an isolated virtual environment or disposable machine if you are unsure. 5) The registry metadata and the SKILL.md disagree about required binaries/config paths; ask the publisher to clarify or review the SKILL.md before trusting automated installation.

Qwen 3.5 — Alibaba's Latest LLM on Your Local Fleet

Qwen 3.5 is the newest and most capable model in the Qwen family. It rivals GPT-4o and Claude 3.5 Sonnet on reasoning, coding, and multilingual benchmarks — and you can run it locally on your own hardware for free.

Supported Qwen models

Model	Parameters	Ollama name	Best for
Qwen 3.5	72B	qwen3.5	Frontier reasoning — rivals GPT-4o
Qwen 3.5	32B	qwen3.5:32b	Strong quality at lower resource cost
Qwen 3.5	14B	qwen3.5:14b	Good balance for mid-range hardware
Qwen 3.5	7B	qwen3.5:7b	Fast on low-RAM devices
Qwen3-Coder	32B	qwen3-coder:32b	Code generation — 80+ languages
Qwen2.5-Coder	7B, 32B	qwen2.5-coder:32b	Proven code model
Qwen3-ASR	—	qwen3-asr	Speech-to-text transcription

Quick start

pip install ollama-herd    # PyPI: https://pypi.org/project/ollama-herd/
herd                       # start the router (port 11435)
herd-node                  # run on each device — finds the router automatically

No models are downloaded during installation. Models are pulled on demand. All pulls require user confirmation.

Use Qwen 3.5 through the fleet

OpenAI SDK

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11435/v1", api_key="not-needed")

# Qwen 3.5 for complex reasoning
response = client.chat.completions.create(
    model="qwen3.5",
    messages=[{"role": "user", "content": "Compare microservices vs monolith architectures"}],
    stream=True,
)
for chunk in response:
    print(chunk.choices[0].delta.content or "", end="")

Qwen3-Coder for code

response = client.chat.completions.create(
    model="qwen3-coder:32b",
    messages=[{"role": "user", "content": "Write a thread-safe connection pool in Go"}],
)
print(response.choices[0].message.content)

Ollama API

# Qwen 3.5 chat
curl http://localhost:11435/api/chat -d '{
  "model": "qwen3.5",
  "messages": [{"role": "user", "content": "Explain attention mechanisms"}],
  "stream": false
}'

Qwen3-ASR speech-to-text

curl http://localhost:11435/api/transcribe \
  -F "file=@meeting.wav" \
  -F "model=qwen3-asr"

Hardware recommendations

Device	RAM	Best Qwen model
Mac Mini (16GB)	16GB	qwen3.5:7b
Mac Mini (32GB)	32GB	qwen3.5:14b or qwen2.5-coder:32b
MacBook Pro (64GB)	64GB	qwen3.5:32b or qwen3-coder:32b
Mac Studio (128GB)	128GB	qwen3.5 (72B) — full quality
Mac Studio (256GB)	256GB	qwen3.5 + qwen3-coder:32b simultaneously

Why Qwen 3.5 locally

• GPT-4o quality — Qwen 3.5 72B matches GPT-4o on MMLU, HumanEval, and MT-Bench
• Zero cost — no per-token charges after hardware
• Privacy — all data stays on your network
• No rate limits — Qwen's cloud API throttles during peak hours. Your hardware doesn't.
• Fleet routing — multiple machines share the load

Also available on this fleet

Other LLMs

Llama 3.3, DeepSeek-V3, DeepSeek-R1, Phi 4, Mistral, Gemma 3, Codestral — same endpoint.

Image generation

curl -o image.png http://localhost:11435/api/generate-image \
  -d '{"model": "z-image-turbo", "prompt": "an AI assistant helping with code", "width": 1024, "height": 1024}'

Embeddings

curl http://localhost:11435/api/embed \
  -d '{"model": "nomic-embed-text", "input": "Qwen 3.5 large language model"}'

Monitor

curl -s http://localhost:11435/fleet/status | python3 -m json.tool
curl -s http://localhost:11435/dashboard/api/health | python3 -m json.tool

Dashboard at http://localhost:11435/dashboard.

Full documentation

• Agent Setup Guide
• API Reference

Contribute

Ollama Herd is open source (MIT):

• Star on GitHub — help others run Qwen locally
• Open an issue — share your Qwen setup, report bugs
• PRs welcome — CLAUDE.md gives AI agents full context. 412 tests, async Python.

Guardrails

• Model downloads require explicit user confirmation — Qwen models range from 4GB (7B) to 42GB (72B).
• Model deletion requires explicit user confirmation.
• Never delete or modify files in ~/.fleet-manager/.
• No models are downloaded automatically — all pulls are user-initiated or require opt-in.