How to Run Qwen3.8-Flash-Next Locally with Ollama (2026 Guide)

What Qwen3.8-Flash-Next Is and Which Tag to Run

Qwen3.8-Flash-Next trades Qwen3.8's dense design for a sparse mixture-of-experts one. Of 125 billion total parameters, only 6 billion activate for any given token, spread across 512 experts, 10 routed per token plus 1 shared expert that always fires. That's why Alibaba can post strong benchmark numbers at this scale while keeping per-token compute closer to a much smaller dense model. The catch is that the full 125B of weights still has to sit in memory regardless of how many activate per step.

The hybrid attention stack interleaves Gated DeltaNet (a linear-attention mechanism that scales more efficiently across long sequences) with Qwen Sparse Attention at a 3:1 ratio, layered under the native 262,144 token context window, extended to 1 million tokens via YaRN. On top of that sits the 51 billion parameter n-gram embedding component, a 20-million-entry table keyed on bigrams and trigrams with gated residual connections, aimed at cutting latency on long agentic sessions specifically, not single-shot prompts.

On Ollama's library page, every currently listed tag is MLX format:

There's no q4_K_M or q8_0-style GGUF build the way Qwen3.8 has. Every path into this model through Ollama runs through Apple's MLX framework. Linux, Windows, and Nvidia GPU setups can't run it locally through Ollama today, full stop, and neither can a rented cloud GPU, since MLX only runs on Apple Silicon.

Install Ollama and Run Qwen3.8-Flash-Next on Apple Silicon

Because the smallest tag is 113GB, plan for a slow first pull even on a fast connection, and check your Mac's unified memory before you start.

Step 1: Confirm Your Mac Has Enough Unified Memory

113GB of weights need real headroom for context and activations on top of the raw download. A Mac Studio or Mac Pro with 128GB unified memory is the realistic minimum for the default tag. 192GB or 256GB gives more room for longer contexts. Even the highest MacBook Pro configuration tops out at 128GB, so this is desktop-class Apple Silicon territory, not laptop territory.

Step 2: Install Ollama

curl -fsSL https://ollama.com/install.sh | sh

Confirm the install and check the version, since MLX-format models need a recent Ollama release:

ollama --version

Step 3: Pull and Run the Default Tag

ollama run qwen3.8-flash-next

Expected output on first run:

pulling manifest
pulling 4f1a9c22...  100% ▕████████████████▏  113 GB
pulling tokenizer...   100% ▕████████████████▏ 4.3 MB
success
>>> Send a message (/? for help)

Step 4: Send a Test Prompt

>>> Write a function that deduplicates a list while preserving order, then explain the tradeoffs of your approach.

Thinking mode is on by default, so the terminal shows reasoning before the final answer. The first response takes noticeably longer than later ones while the full model loads into unified memory.

Step 5: Verify the Download

ollama list

`qwen3.8-flash-next:latest` should show at 113 GB. That confirms the full MLX weights actually landed on disk, not just a placeholder or manifest-only cloud entry.

Vision Input, Tunable Reasoning, and Agentic Tool Use

Flash-Next accepts image input alongside text. Alibaba calls out STEM diagrams, charts, documents, hour-scale video analysis, and computer-use or mobile-use tasks specifically, a step past general photo description.

Send an Image from the Terminal

ollama run qwen3.8-flash-next "What does this architecture diagram show? ./diagram.png"

Control Reasoning Effort via the API

curl http://localhost:11434/api/chat -d '{
  "model": "qwen3.8-flash-next",
  "messages": [
    { "role": "user", "content": "Plan a migration from a monolith to microservices for this codebase." }
  ],
  "think": true,
  "reasoning_effort": "xhigh",
  "stream": false
}'

`reasoning_effort` takes low, medium, or xhigh, a higher ceiling than Qwen3.8's low/medium/high scale. Flash-Next is tuned for harder agentic and coding benchmarks, and that extra tier gives it room to actually use it. Lower settings trade reasoning depth for speed on simpler prompts.

Use Qwen3.8-Flash-Next in Scripts and Coding Agents

Python Example

from ollama import chat

response = chat(
    model="qwen3.8-flash-next",
    messages=[{"role": "user", "content": "Review this pull request diff for edge cases."}],
    think=True,
)
print(response["message"]["content"])

OpenAI-Compatible Endpoint for Agent Tools

model:
  default: qwen3.8-flash-next
  provider: custom
  base_url: http://localhost:11434/v1
  context_length: 262144

Point any agent tool that speaks the OpenAI-compatible API, Hermes Agent, OpenClaw, and similar, at this config to use Flash-Next's full 256K native context. The n-gram embedding layer's latency work targets exactly this kind of long, tool-heavy agent session, not one-off chat prompts. This is where the architecture's design choices actually show up in day-to-day use.

Troubleshooting

Frequently Asked Questions