Mistral-7B-NovaKV

NovaKV compresses the KV cache of a frozen Mistral-7B-Instruct-v0.3 by a post-hoc low-rank factorisation of the K and V projections. The cache is 45.4% smaller, measured as a real allocation delta rather than estimated from ranks. The base model is not retrained.

Code, full results and reproduction commands: https://github.com/declercqcharles/NovaKV-public

This is a delta, not a full model

novakv_mistral7b_v03_joint_delta.pt (2.45 GB) carries only the tensors that differ from mistralai/Mistral-7B-Instruct-v0.3 — the compressed K/V factors and the q/o projections. Every other weight was verified bit-identical to the base model and is not redistributed here. It cannot be loaded with from_pretrained; it needs the loader from the repository above, which fetches the base model and applies the delta on top.

git clone https://github.com/declercqcharles/NovaKV-public
cd NovaKV-public && pip install -r requirements.txt
hf download declercqcharles/Mistral-7B-NovaKV novakv_mistral7b_v03_joint_delta.pt --local-dir .

PYTHONPATH=. python scripts/eval.py --model mistralai/Mistral-7B-Instruct-v0.3 \
  --weights novakv_mistral7b_v03_joint_delta.pt \
  --ppl --ppl-datasets wikitext2,c4 --ppl-seqlen 2048

Results

┌──────────────────────┬────────┬────────┐
│                      │ Dense  │ NovaKV │
├──────────────────────┼────────┼────────┤
│ PPL WikiText-2 @2048 │   5.50 │   5.76 │
├──────────────────────┼────────┼────────┤
│ PPL C4 @2048         │   8.85 │   9.47 │
├──────────────────────┼────────┼────────┤
│ RULER                │ 97.12% │ 96.35% │
├──────────────────────┼────────┼────────┤
│ LongBench            │ 43.44% │ 39.72% │
├──────────────────────┼────────┼────────┤
│ MMLU                 │ 59.98% │ 57.44% │
├──────────────────────┼────────┼────────┤
│ HumanEval-instruct   │ 42.07% │ 32.93% │
├──────────────────────┼────────┼────────┤
│ MBPP-instruct        │ 42.60% │ 35.20% │
├──────────────────────┼────────┼────────┤
│ Cache, KB/token      │ 131.07 │  71.51 │
└──────────────────────┴────────┴────────┘

Code generation is where the cost concentrates: HumanEval and MBPP lose 7 to 9 points, while
perplexity, long-context retrieval and general knowledge stay close to dense.

Single-stream decode latency, fixed-shape cache with a captured CUDA graph, against the stock
Hugging Face dense path: 10.29 ms/token at context 256, 12.50 at 2048, 18.43 at 8192, versus a
flat ~27 ms for dense. The dense baseline is not itself CUDA-graph captured. Above batch 1 the
compressed path is slower than dense.

RULER is averaged over 12 of its 13 subtasks; ruler_qa_hotpot fetches its data from a host that
has been unreachable since July 2026, and is excluded from the dense baseline identically.

Mistral's optional sliding-window attention is not implemented in this pipeline. It is harmless
with sliding_window: null, which is the case for this checkpoint, but would be silently wrong
otherwise.

License

Apache 2.0, as the base model. The code in the linked repository is MIT.