Skip to content
Language models

Qwen3.8-Flash-Next Gyro-Coder

a 125B coding model on a single 24 GB GPU

3 min readupdated

Qwen3.8-Flash-Next Gyro-Coder is a specialized GGUF build of the massive Qwen/Qwen3.8-Flash-Next language model (125B parameters, MoE — mixture-of-experts architecture) by the agentionai team. The developers pruned half of the experts, keeping only the 256 out of 512 per layer dedicated to programming, and compressed the weights with a custom trellis coder down to 2.125–3.2 bits per weight. As a result, the model fits onto a single consumer GPU like an RTX 3090, 4090, or 7900 XTX while maintaining near-identical token predictions to the original on coding tasks. It is built for developers who want a heavy-hitting local assistant for writing, refactoring, and debugging code.

Key features

  • Smart expert pruning: evolutionary selection leaves only the 65B active parameters critical for programming in GPU memory.
  • Two variants for different VRAM capacities: the base TQ2_0 takes up 19.92 GiB (fitting into 24 GB even with a 64k context), while Gyro-Coder-L (TQ3_0) requires 27.54 GiB for 32 GB cards.
  • Reliable tool calls (v1.2): the updated release restored three key experts, completely eliminating broken tags during tool calls in agentic sessions.
  • Lightweight downloads (-noPLE): the n-gram table (PLE) can be downloaded separately once, shrinking the model file size from 75.8 GB down to 21.4 GB.

What you can do

  • Hook up to an IDE or agentic framework: spin up a local OpenAI-compatible API and point your editor plugin to it for autocomplete and multi-file edits.
  • Automate refactoring and bug fixing: feed the model stack traces, SQL queries, or C++, Python, TypeScript, and Rust code with clear instructions.
  • Run inside coding agents with tool calls: Gyro-Coder-L v1.2 completed 292 tool calls across agentic tasks without a single formatting error.

Example

hf download agentionai/Qwen3.8-Flash-Next-Gyro-Coder-GGUF Qwen3.8-Flash-Next-Gyro-Coder-TQ2_0.gguf --local-dir ~/models
docker run --rm -it --gpus all -e NVIDIA_DRIVER_CAPABILITIES=all \
  -v ~/models:/models -p 8080:8080 agentionai/agention-llama:server \
  -m /models/Qwen3.8-Flash-Next-Gyro-Coder-TQ2_0.gguf -ngl 999 -fa on --jinja --ngram-on-disk -c 65536
Read moreDetails
Quality: top-1 agreement with the full Q8_0 model reaches 94.5% on C++ and 93.1% on Python (95.5% and 94.0% for Gyro-Coder-L); passes 13 out of 16 automated benchmark tasks.
Hardware and requirements: the 19.92 GiB weights (21.1 GiB with a 64k context in q8_0) require a 24 GB VRAM GPU and a fast NVMe SSD for offloading the n-gram table (--ngram-on-disk).
Limitations: stock llama.cpp, Ollama, and LM Studio do not support this format yet — you need the fork from agentionai or strix-llama.cpp; general knowledge and character tables (such as the Greek alphabet) are degraded due to pruning half the experts.
Discuss 89 64.7K

Will it run?

Won't run

Not enough memory even for 4-bit. Below are GPUs you can rent.

  • GGUF TQ3_0weights 106 GB · ready files117 GB
  • GGUF TQ2_0weights 91 GB · ready files100 GB

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: Won't fit any home GPU — renting is cheaper.

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # agentionai/Qwen3.8-Flash-Next-Gyro-Coder-GGUF
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download agentionai/Qwen3.8-Flash-Next-Gyro-Coder-GGUF --local-dir ~/models/Qwen3.8-Flash-Next-Gyro-Coder-GGUF
    # The model is in ~/models. How to run it is in the model card (link above).
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

  • MTP (multi-token prediction)0.34.0from November 16, 2026new release

    1.5–2× generation speed

    If the model has MTP layers (DeepSeek-V3/R1, Qwen3-Next, GLM, MiMo and others), it predicts several tokens per step itself.

    How to enable: vLLM / SGLang: --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. Check the model card.

    Earlier generations (2)
    • EAGLE-3 — Speculative decoding with a small draft head. For models without their own MTP layers.
    • EAGLE-2 / Medusa — Earlier draft-head generations. Still work if that is all you find.
  • FP8 KV cache0.34.0from November 16, 2026new release

    2× more context in the same memory

    Keeps the attention cache in 8 bits, so long contexts fit where they did not.

    How to enable: vLLM: --kv-cache-dtype fp8; llama.cpp: -ctk q8_0 -ctv q8_0.

  • vLLM / SGLang0.34.0from November 16, 2026new release

    many times more tokens/s on a server

    Server engines: batching, FlashAttention, FP8 and speculative decoding out of the box.

    How to enable: vllm serve <model> — an OpenAI-compatible API.

    Earlier generations (1)
    • llama.cpp — GGUF models on any GPU, Mac or even CPU; some layers can stay in RAM.

Model family

New posts about versions of Qwen3.8-Flash-Next (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 12All on Hugging Face
Fine-tunes 8All on Hugging Face

About the model

agentionai/Qwen3.8-Flash-Next-Gyro-Coder-GGUF
Author
agentionai
Category
Language models
Released
Posted

On NeyroLearn

Hugging Face

64.7K89

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments