Skip to content
Other

Qwen3.8-27b-Terse-Coder-Basalt

a compact version of Qwen3.8-27B for fast coding

2 min readupdated

A ready-to-run package of Qwen3.8-27B by Shockem, fine-tuned for coding and packaged in the Basalt format (a file format for running quantized models locally). The main highlight of the Terse-Coder version is cutting the chain-of-thought roughly tenfold while maintaining answer accuracy. The model is based on Qwen/Qwen3.8-27B with the Terse-Coder adapter merged in and quantized in mixed precision (NVFP4 and FP8), bringing the file size down to 21.6 GB. It suits developers who need a local 27B-parameter model that doesn't waste extra tokens and time on lengthy reasoning before outputting code.

Key features

  • Concise reasoning: uses roughly 10x fewer tokens for chain-of-thought on coding tasks without sacrificing solution accuracy.
  • Mixed quantization (21.6 GB): MLP layers and the output layer are compressed to NVFP4, attention projections and GDN to FP8 E4M3, while embeddings remain in BF16.
  • Vision and long-context support: the package includes a vision projector, and the model was tested at a 131,072-token context length with an int8 KV cache.
  • Tool use: passed an internal benchmark suite (15/15 in 64 seconds) covering instruction following, JSON output, tool calling, and code generation.

What you can do

  • Local coding assistant: hook it up to an IDE or agentic pipeline (for agentic mode, the author recommends the Shockem/froggeric-terse-coder template) for fast code generation with minimal token overhead.
  • Run on dual mid-range GPUs: deploy it across two 16 GB VRAM cards (the model card includes benchmarks on dual RTX 5060 Ti 16 GB GPUs with layer splitting).
  • Automation and function calling: use it in local scripts that require strict JSON output, tool calling, and up to 128k context.

Example

basalt serve Qwen3.8-27B-Terse-Coder-NVFP4.basalt --spec 0 --kv int8
Read moreDetails
Original checkpoint metrics on vLLM (not re-benchmarked in the Basalt format): HE+ 91.5, MBPP+ 79.4, GSM8K 98.5, GPQA 79.8.
Speed benchmarks on Basalt 0.4.1 (2× RTX 5060 Ti 16 GB, 131,072 context, int8 KV): prefill — 3,491–5,373 tok/s (across prompt lengths from 512 to 64k), generation — 22.1 tok/s.
Limitation: when splitting layers across two GPUs, you must pass the --spec 0 flag (with spec >= 2, the built-in FP8 MTP drafter crashes with a memory error on the first token).
Recommended sampling parameters (baked into the package): temperature 0.6, top_p 0.95, top_k 20. License: Apache 2.0.
Discuss Telegram 0 20

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

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.

    # Shockem/Qwen3.8-27b-Terse-Coder-Basalt
    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 Shockem/Qwen3.8-27b-Terse-Coder-Basalt --local-dir ~/models/Qwen3.8-27b-Terse-Coder-Basalt
    # 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.

Model family

New posts about versions of Qwen3.8-27b-Terse-Coder-Basalt (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

Shockem/Qwen3.8-27b-Terse-Coder-Basalt
Author
Shockem
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

200

Ask AI about this model

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

Comments

Computer vision

HITCH

HITCH