alphamike/kvbridge-ministral3-3b-to-8b

KVBridge Ministral 3 3B → 8B Mapper

This repository releases a fitted cross-model KV-cache mapper from mistralai/Ministral-3-3B-Reasoning-2512 to mistralai/Ministral-3-8B-Reasoning-2512. It is an independent reproduction of the method in NVIDIA's paper Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse, implemented against the authors' KVBridge repository.

This is a general-purpose replication artifact and is not affiliated with or endorsed by NVIDIA or Mistral AI.

What this artifact does

KVBridge allows a smaller model to prefill a context and maps its KV cache into the cache space expected by a larger model in the same family. The larger model can then continue from that mapped cache without recomputing the complete context itself.

This repository is not a standalone language model. It contains the linear mapper used between the two exact model revisions below; users must separately obtain both base models under their respective licenses.

Role	Model	Pinned revision
Source/prefill	mistralai/Ministral-3-3B-Reasoning-2512	4a36357c811bf511a7b625d132e12f22408aac91
Target/decode	mistralai/Ministral-3-8B-Reasoning-2512	81eaece1948f3875421d9a45bc55487d10e2d894

Do not assume compatibility with other checkpoints, revisions, KV geometries, or quantized model variants without refitting and validation.

Files

File	Purpose
mapper.safetensors	Fitted per-target-layer key/value affine maps (float32, 7.42 GB)
mapper.json	Model pair, RoPE, layer selection, fit, and format metadata
kvbridge_ministral3_v1.patch	General Ministral-3 text-decoder and exact-YaRN compatibility patch for KVBridge
example.py	Minimal cache-transfer example
artifact_manifest.json	Machine-readable provenance and evaluation summary
checksums.sha256	SHA-256 integrity checks

The weights are preserved in float32 exactly as fitted. For runtime use, load them as BF16 to reduce resident mapper memory:

import torch
from kvbridge import Mapper

mapper = Mapper.from_pretrained(
    "alphamike/kvbridge-ministral3-3b-to-8b",
    device="cuda:0",
    dtype=torch.bfloat16,
)

Install the compatible KVBridge revision

The upstream KVBridge code did not directly load the text decoder inside the multimodal Ministral-3 repositories at the pinned revision used here. Apply the included general compatibility patch:

git clone https://github.com/spotta85/kvbridge.git
cd kvbridge
git checkout fdc11d31f005d7d4cec949b72e9032f2093dd6b8
wget https://huggingface.co/alphamike/kvbridge-ministral3-3b-to-8b/resolve/main/kvbridge_ministral3_v1.patch
git apply kvbridge_ministral3_v1.patch
pip install -e .

The patch adds text-only loading for Ministral-3 checkpoints and preserves their exact YaRN rotary-position specification during cache transfer. See example.py for a small end-to-end cache handoff. The current transfer runtime supports batch size 1.

Mapper fit

The mapper follows the paper's content-space, closed-form ridge procedure:

• 500 calibration texts × 1,024 tokens;
• stride 4, giving 128,000 token observations;
• all 26 source layers used for each of 34 target layers (top_k=26);
• ridge coefficient 0.01;
• 8 KV heads with head dimension 128;
• fitted key/value parameters: 1.85B;
• fitting time: 200.4 seconds;
• peak fitting GPU memory: 16.08 GiB;
• mean fit R²: 0.8789 for keys and 0.7611 for values.

The R² values are fit diagnostics, not downstream task accuracy. The calibration corpus content hash is 985f7bc5a29212748fc05de6f4cdebe9a024f47e2d6c5c7eae084d80cd690ead.

Reproduction accuracy

These are frozen, full-task results from this independent Ministral 3B→8B replication. “Retention” is mapped accuracy divided by native target accuracy. The paper values are included only as a reference; the model pair and execution environment are not identical to every experiment in the paper.

Task	Native 8B	Mapped 3B→8B	Raw retention	Floor-normalized retention
ARC-Challenge	60.15%	57.34%	95.32%	91.99%
HellaSwag	76.86%	74.13%	96.45%	94.74%
WinoGrande	71.43%	67.01%	93.81%	79.37%
MMLU, 5-shot	70.15%	58.96%	84.05%	75.22%
GSM8K, 8-shot CoT, strict match	82.94%	19.94%	24.04%	24.04%
Mean	—	—	78.73%	73.07%

GSM8K flexible answer extraction gives 73.01% mapped accuracy, but the strict exact-match score above is the primary reported result. The large gap between the two makes GSM8K particularly sensitive to answer formatting and evaluation protocol.

For context, the paper's five-task reference values for its reported native and mapped setup were 76.2% mean raw retention and 65.9% mean floor-normalized retention. This artifact's result should be read as a reproduction on the pinned Ministral pair, not as an exact hardware- or model-matched rerun of every paper result.

Prefill and mapping latency

Measured on one NVIDIA A100 40 GB GPU in BF16, batch size 1, with 2 warmups and 7 timed trials. “Warm map” is mapper execution after the source KV cache already exists, matching the paper's cache-transfer timing boundary. “Cold source+map” includes the 3B source prefill and is the relevant end-to-end replacement for native 8B prefill.

Context tokens	Native 8B prefill	Source 3B prefill	Warm map	Cold source+map	Warm speedup	Cold speedup
64	28.8 ms	22.0 ms	24.4 ms	46.4 ms	1.18×	0.62×
512	51.5 ms	26.1 ms	35.5 ms	61.6 ms	1.45×	0.84×
1,024	91.7 ms	42.1 ms	50.2 ms	92.4 ms	1.83×	0.99×
2,048	170.6 ms	82.6 ms	80.0 ms	162.6 ms	2.13×	1.05×
4,096	336.9 ms	167.1 ms	141.8 ms	308.9 ms	2.38×	1.09×
8,192	706.4 ms	361.5 ms	275.5 ms	637.0 ms	2.56×	1.11×
16,384	1,564.5 ms	837.9 ms	542.2 ms	1,380.1 ms	2.89×	1.13×

Absolute latency depends strongly on hardware, kernels, dtype, software stack, and whether source prefill is included. A simultaneous 3B+8B 32K run did not fit on the single 40 GB A100 used here. The paper's 32K timing used 8×H100 NVLink and excluded source prefill, so those absolute numbers are not directly comparable.

Limitations

• Accuracy is not guaranteed to match native 8B prefill; degradation is highly task dependent, and strict GSM8K is a clear failure case.
• This artifact is tied to the exact source/target revisions listed above.
• Transfer currently supports batch size 1.
• The 7.42 GB float32 mapper is large; BF16 runtime loading is recommended.
• The base models must be downloaded separately and may require substantial GPU memory. Follow their licenses and acceptable-use terms.
• This artifact is for research and engineering evaluation. Validate it for the intended application before deployment.

Integrity

2de7fbd8d88ebe12433b73878ce5f847b2b3fa32b177b98f2acd26e0ba2489c3  mapper.json
dd37580166c40dc86afaccf6e80829b44af29a033e1e3675c85a1d28eff64093  mapper.safetensors
fd0f7e734fabdcfee4585b338a9168d62159fba39015aae077aa1ed7c5b253e0  kvbridge_ministral3_v1.patch

Citation

If this artifact is useful, cite the original method:

@article{heo2026crossmodel,
  title={Cross-Model KV Cache Transfer in LLM Families: A Closed-Form Linear Mapping for Prefill Reuse},
  author={Heo, Taekyung and Shafipour, Rasoul and Zhao, Ritchie and Golub, Maximilian and Kamani, Mohammad Mahdi and Borkar, Ritika and Chandran, Makesh Tarun and Zardoshti, Pantea and Darvish Rouhani, Bita},
  journal={arXiv preprint arXiv:2608.03893},
  year={2026}
}

Licenses and acknowledgements

The mapper artifact and compatibility patch are released under MIT. The pinned Mistral base-model repositories are licensed separately under Apache-2.0. This release builds on the method and open-source implementation from the NVIDIA paper authors; please credit their work and review the upstream repositories.