FredQuant/corriente-qwen38-flash-next

Qwen3.8-Flash-Next — Milked Edition (Q4_K_M, 16GB)

The full 82GB model, distilled into a single 16GB file that runs anywhere. We took Qwen3.8-Flash-Next — Alibaba's 125B-parameter Mixture-of-Experts model — and extracted its core intelligence into a single, deployment-ready GGUF. Same brain. 80% smaller. Runs on a laptop.

Why This Exists

The original model ships as 82GB across multiple shard files. That's fine for data centers, but most teams can't deploy that. We extracted the backbone — the 6B active parameters that actually think — and requantized it to Q4_K_M. The result: a single 16GB file that fits in 20GB of RAM. What we kept: Full reasoning capability, 262K context, all 512 experts. What we removed: The 24GB N-gram embedding lookup table (training-only, not needed for inference).

The Numbers

Property	Value
Total parameters	125B
Active parameters	6B per token
Experts	512 (10 routed + 1 shared)
Quantization	Q4_K_M
File size	16GB
Context length	262,144 tokens
Architecture	Gated DeltaNet + Qwen Sparse Attention
Original size	82GB
Size reduction	80%
License	Apache 2.0

Performance

• Beats Claude Opus 4.6 on SWE-bench Pro (62.5 vs 53.4)
• 12× cheaper than Qwen3.8-Max ($0.16/$0.47 per M tokens)
• 4.8% active compute — like running a 6B model with 125B intelligence
• Runs on any machine with 20GB+ RAM — no GPU required

Quick Start

Ollama (easiest)

# Create the model
cat > Modelfile << 'EOF'
FROM ./qwen3.8-flash-next-Q4.gguf
PARAMETER num_ctx 16384
EOF
ollama create qwen38-flash -f Modelfile
ollama run qwen38-flash

llama.cpp (server)

./llama-server \
-m qwen3.8-flash-next-Q4.gguf \
-c 16384 \
--host 0.0.0.0 \
--port 8080

Python

from llama_cpp import Llama
llm = Llama(model_path="./qwen3.8-flash-next-Q4.gguf", n_ctx=16384)
output = llm("Explain quantum computing in simple terms", max_tokens=200)
print(output["choices"][0]["text"])

Reasoning & Tool Use (important for agentic apps)

This model is a reasoning model: before answering, it runs an internal thinking pass (same style as DeepSeek-style chain-of-thought). How you see it depends on the runtime:

• Ollama (recommended): reasoning is natively separated — content contains only the final answer, and the thinking pass comes back in its own thinking field. Ollama cannot bake the toggle into the model (PARAMETER think is unsupported), so pass it per request:

  curl http://localhost:11434/api/chat -d '{
    "model": "qwen38-flash",
    "think": false,
    "messages": [{"role": "user", "content": "Hello"}]
  }'
  

  Set "think": false for tool-calling / agentic loops and chat UIs where you want the final answer only. Models that don't reason simply ignore the flag.
• llama.cpp / raw GGUF: reasoning is emitted as plain text. In agentic or chat apps, strip the thinking delimiters or request short, direct answers via the system prompt.
• Verified on this file (Sep 2026): with think on, the answer stream stays clean and the reasoning is reported separately; with think off, replies are fully direct. Native tool calls work correctly either way.

What We Do

We don't just share models. We make them work for you. This GGUF is one example of what we do at Corriente. We take powerful open-source models and make them deployable, customizable, and aligned to your specific work.

Our Services

Model Distillation & Quantization We extract the intelligence from large models and compress them for your hardware. From 82GB to 16GB. From cloud-only to run-anywhere. We've done this with Qwen, Llama, DeepSeek, and more. Custom Model Training We fine-tune models on your data, your domain, your terminology. Legal firms, hospitals, engineering teams — we build models that speak your language. Not generic AI. Your AI. Fleet Deployment We deploy models across multi-node clusters for parallel inference. Our infrastructure runs on NVIDIA DGX Sparks — enterprise-grade hardware at a fraction of the cost. We handle the orchestration so you don't have to. Data Cleaning & Preparation Garbage in, garbage out. We clean, structure, and prepare your data for training. Our pipeline handles deduplication, quality filtering, domain tagging, and format normalization. We've processed millions of records across dozens of domains. Integration & Support We don't drop a model on your doorstep and walk away. We integrate it into your workflow, train your team, and provide ongoing support. We're builders, not vendors.

Who We've Built For

• Research teams needing local inference without API costs
• Organizations with sensitive data that can't leave their network
• Teams that need models trained on domain-specific knowledge
• Companies tired of paying per-token for generic AI

How to Work With Us

If you need a model that thinks about your work, not just general knowledge — we can build it. Email: fredramirez3rd@gmail.com Website: corriente.ai We leave no one behind.

Credits

• Original model: Alibaba Qwen Team
• GGUF conversion: Unsloth
• Milked & deployed: Corriente LLC

Show Your Support ⭐

This model is free because we believe powerful AI should be in everyone's hands — but a small star goes a long way.

If this model helped you, we'd truly appreciate a ⭐ on this repo. It takes one click, it's free, and it tells us (and the world) this work matters. It also helps more people find a model they can actually run.

And if you want to go further — tell us what you built with it, or share it with a team that needs local AI. Word of mouth from people who actually use the work is the best fuel we know.

License

Apache 2.0 — same as the original. Use it, share it, build on it.

Built by Corriente — the future is frequency. ⭐ if you agree.