ministral-3-3b-reasoning-2512-gguf
Consumer-supplied Wekala model-library assets derived from the attributed official publisher. Models and this repository are distributed separately from the Wekala SDK.
Origin and licence
| Field | Value |
|---|---|
| Original developer | Mistral AI |
| Original model | mistralai/Ministral-3-3B-Reasoning-2512 |
| Pinned source revision | 4a36357c811bf511a7b625d132e12f22408aac91 |
| Original publication | 2025-12-02 (day) |
| Date evidence | Publisher source |
| Licence | apache-2.0; preserved in LICENSE |
| Full source parameters | 4,251,743,232 (includes embeddings and source modalities) |
| Nile Academy role | Official artifact mirror or explicitly recorded format conversion; no additional training |
Original publisher cards and available notices are preserved in provenance/. The original model licence governs the weights; this repository grants no rights to the Wekala SDK.
Files and quantization
Q8_0 is the higher-precision comparison; Q6_K/Q5_K_M trade memory for precision; Q4_K_M needs paired task evaluation. No general quality-loss or 8GB-fit qualification yet.
| File | Quantization | Size (GB) | Origin |
|---|---|---|---|
| Ministral-3-3B-Reasoning-2512-BF16-mmproj.gguf | BF16 PROJECTOR | 0.842 | official publisher export |
| Ministral-3-3B-Reasoning-2512-Q4_K_M.gguf | Q4_K_M | 2.147 | official publisher export |
| Ministral-3-3B-Reasoning-2512-Q5_K_M.gguf | Q5_K_M | 2.474 | official publisher export |
| Ministral-3-3B-Reasoning-2512-Q8_0.gguf | Q8_0 | 3.652 | official publisher export |
Exact file SHA-256 values, upstream paths/revisions and conversion receipts are recorded in model-metadata.json.
Published benchmarks
Publisher-reported source-model scores follow. They are not Wekala measurements of these quantized files. Harness, precision, reasoning settings and task versions differ; do not treat them as a normalized cross-model ranking.
| Benchmark | Score (%) | Mode / settings |
|---|---|---|
| AIME25 | 72.1 | reasoning; Publisher source protocol; not normalized across models or measured on this quantization. |
| GPQA Diamond | 53.4 | reasoning; Publisher source protocol; not normalized across models or measured on this quantization. |
| LiveCodeBench (version unspecified) | 54.8 | reasoning; Publisher source protocol; not normalized across models or measured on this quantization. |
Qualification and use
Downloadable declarative asset. Wekala native chat/tool/template, sustained GPU fit and mobile/browser qualification are pending; no automatic backend installation.
Wekala broad quantization-quality qualification: not qualified. No general “no significant quality loss” claim is made. Official QAT or publisher comparisons are attributed separately. Before recommendation, use paired reference/variant evaluation across instruction, knowledge, math, multilingual/Arabic and tool tasks under the registered Wekala campaign.
Device fit is separate: weight size excludes context state, scratch buffers and driver headroom. CPU and CUDA are execution backends for the same weight file, not separate models. Mobile/browser and sustained 8GB-VRAM operation require their own measured profiles.
Models can produce incorrect or unsafe outputs. Tool calls are proposals subject to the consuming application’s permissions and controller admission.
Original-precision checkpoint
The original BF16 checkpoint is preserved under original/4a36357c811bf511a7b625d132e12f22408aac91/, copied from mistralai/Ministral-3-3B-Reasoning-2512 at the exact pinned revision. Original tensor files total 7,698,241,056 bytes. This includes the source modalities; publisher executable Python is excluded. Source licences, original release date and published benchmarks above still apply.
| Download | Precision | Bytes | Scope |
|---|---|---|---|
| model-original-bf16.gguf | BF16 original precision | 6,866,743,968 | Text-only GGUF reference; separate from the complete original checkpoint |
Original-precision options have no 8 GB eligibility ceiling. File size is not total RAM/VRAM usage. These are not Q8 and are not dequantized QAT/Q4 weights. GGUF conversion fidelity, runtime profiles and sustained device fit remain separate qualification gates; no lossless-conversion or quality guarantee is implied. See the immutable source and reference record.