Qwen3.5 0.8B Weight-Space Programming — Artemis

ML-Intern-lab’s edited Qwen checkpoint as a 532 MB text-only GGUF, down from 1.517 GB for the BF16 text conversion: about 65% smaller. Compressed with Artemis Adaptive Compression by TurinTech.

Size and fidelity

Agreement means the same next token as the edited BF16 parent, not maths accuracy. Lower KL means closer output probabilities. Each panel has 15,300 scored positions (60 chunks of 512 tokens).

Model type: chat. Recommended file: Qwen3.5-0.8B-weight-space-programming-Artemis-Q4_K_M.gguf.

Run with Hugging Face

Verification status: verified.

• Download the file:

  hf download turintech/Qwen3.5-0.8B-weight-space-programming-Artemis Qwen3.5-0.8B-weight-space-programming-Artemis-Q4_K_M.gguf --local-dir .
  

• Run with a current llama.cpp build (thinking enabled):

  llama-cli -m Qwen3.5-0.8B-weight-space-programming-Artemis-Q4_K_M.gguf \
    -c 34816 -n 32768 --jinja --reasoning on \
    --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0 \
    --presence-penalty 1.5 --repeat-penalty 1 --repeat-last-n 34816 \
    --samplers "penalties;top_k;top_p;temperature" --seed 0 --no-context-shift
  

Files

GB means decimal download size, not runtime memory.

Limitations

• Text only; no vision projector or MTP module is included.
• The upstream checkpoint’s reported MATH-500 score is not a measured score for this GGUF. Quantized long-reasoning quality is still being evaluated.
• Download size is not runtime memory usage.

Benchmark results and methodology

Full confirmation evaluation; not an untouched final evaluation.

Evaluation protocol

Measured on WikiText-2 test and OASST1 validation: 60 chunks of 512 tokens per panel, with 15,300 scored positions per panel. These are fixed historical fidelity panels, not untouched task tests. Native CPU evaluation used a fresh edited-parent reference; KL is approximate because the reference cache stores clipped, quantized log probabilities.

Source

Base model: ML-Intern-lab/Qwen3.5-0.8B-weight-space-programming. License: Apache 2.0 (apache-2.0).

Artifact attribution

• Qwen3.5-0.8B-weight-space-programming-Artemis-Q4_K_M.gguf: Quantized by TurinTech from TurinTech conversion of the pinned ML-Intern-lab checkpoint artifacts. Reused conversion: llama.cpp GGUF conversion by ggml-org.

File identities

Original source: ML-Intern-lab/Qwen3.5-0.8B-weight-space-programming at f4a5dd43dae4a5e42b441a31d692570751b3d92c.