Full training-run archive for Kuza (East Africa agricultural assistant), fine-tuned from unsloth/Qwen3.5-4B. Weights, logs, checkpoints, GGUFs, and provenance are stored with the same layout as $KUZAWORKDIR/kuza-qwen-3.5-4b/.
Model Card
By Kuza AI, published under apache-2.0, revision 9c2d90a8f6a3.
Full training-run archive for Kuza (East Africa agricultural assistant), fine-tuned from unsloth/Qwen3.5-4B. Weights, logs, checkpoints, GGUFs, and provenance are stored with the same layout as $KUZAWORKDIR/kuza-qwen-3.5-4b/. This derivative is subject to the Apache 2.0 license of the Qwen base model. - 100% English train from kuzaai/kuzasftenglish - 35% Swahili from kuzaai/kuzasftswahili - 8% HuggingFaceH4/norobots - 5% adversarial from kuzaai/kuzasftadversarial - all multiturn from kuzaai/kuzasftmultiturn Sequence length 1024, 2 epochs, LR 0.0002. Thinking is off (--reasoning off when llama.cpp supports it). GGUFs are text-only (MTP/nextn kept; vision/audio dropped). ssmout is Q6K on…
Read Kuza AI's full model card
Full training-run archive for Kuza (East Africa agricultural assistant),
fine-tuned from unsloth/Qwen3.5-4B.
Weights, logs, checkpoints, GGUFs, and provenance are stored with the same
layout as $KUZA_WORK_DIR/kuza-qwen-3.5-4b/.
This derivative is subject to the Apache 2.0 license of the Qwen base model.
Training mix
- 100% English train from
kuzaai/kuza_sft_english - 35% Swahili from
kuzaai/kuza_sft_swahili - 8%
HuggingFaceH4/no_robots - 5% adversarial from
kuzaai/kuza_sft_adversarial - all multiturn from
kuzaai/kuza_sft_multiturn
LoRA: RsLoRA r=32, alpha=64, no QAT.
Sequence length 1024, 2 epochs, LR 0.0002.
Thinking is off (--reasoning off when llama.cpp supports it). GGUFs are text-only
(MTP/nextn kept; vision/audio dropped). ssm_out is Q6_K on every candidate.
Files
adapter/— PEFT adapter, tokenizer, SFT metrics and manifeststraining/— Trainer checkpoints includingcheckpoint-bestmerged_bf16/— text-only merged Hugging Face BF16 weightsreference/— text-only BF16 GGUF (kuza-bf16.gguf) and smoke logimatrix/— calibration corpus, eval corpus, imatrix, logsquants/— quantized GGUF candidates:q4_k_m_imatrix/kuza-qwen-q4_k_m.gguf(q4_k_m)q4_k_xl_ssm/kuza-qwen-q4_k_xl.gguf(q4_k_m)q4_k_s_ssm/kuza-qwen-q4_k_s.gguf(q4_k_s)screen/— GPU KLD diagnostic, hidden-set scores, andresults.jsonwinnerprovenance/— copied adapter metrics, recipes, screen JSONupload_manifest.json— path, size, and sha256 for every uploaded file
Download
huggingface-cli download kuzaai/kuza-qwen-3.5-4b --local-dir ./kuza-qwen-3.5-4b
screen/results.json ranks by hidden-set accuracy, then GGUF size, then GPU
TPS. Do not treat GPU TPS as an ADTC laptop measurement.
Identity and Version
- Repository
- kuzaai/kuza-qwen-3.5-4b
- Publisher
- Kuza AI
- Task
- Not stated by the source
- Modality
- Other
- Library
- peft
- Parameters
- Not stated by the source
- Languages
- en, sw
- Revision
- 9c2d90a8f6a34df7e1bfe40eb85e5e655faa5284
- First published
- 2026-09-18
- Last updated
- 2026-09-18
Files and Weights
121 files, 33.2 GB in total. The weights are 33 files totalling 29.0 GB in bin, gguf, pt, pth, safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| adapter/adapter_model.safetensors | Weights | 223.4 MB | acb5d01c2f9a |
| merged_bf16/model-00001-of-00002.safetensors | Weights | 5.0 GB | 8086e3be3757 |
| merged_bf16/model-00002-of-00002.safetensors | Weights | 3.4 GB | e1c20c3d73a0 |
| provenance/adapter_model.safetensors | Weights | 223.4 MB | acb5d01c2f9a |
| quants/q4_k_m_imatrix/kuza-qwen-q4_k_m.gguf | Weights | 2.7 GB | 4a7a8921cb83 |
| quants/q4_k_s_ssm/kuza-qwen-q4_k_s.gguf | Weights | 2.6 GB | a8b7cfe78182 |
| quants/q4_k_xl_ssm/kuza-qwen-q4_k_xl.gguf | Weights | 3.0 GB | 82fa31b2fa8c |
| reference/kuza-bf16.gguf | Weights | 8.4 GB | 8d355227aea4 |
| training/checkpoint-2180/adapter_model.safetensors | Weights | 223.4 MB | 117561a10de2 |
| training/checkpoint-2180/optimizer.pt | Weights | 447.0 MB | 92de09b702f1 |
| training/checkpoint-2180/rng_state.pth | Weights | 14.6 KB | 52e8d3009833 |
| training/checkpoint-2180/scheduler.pt | Weights | 1.5 KB | 774c899a3c80 |
| training/checkpoint-2180/training_args.bin | Weights | 5.8 KB | f102db021651 |
| training/checkpoint-4360/adapter_model.safetensors | Weights | 223.4 MB | fdd033b851ca |
| training/checkpoint-4360/optimizer.pt | Weights | 447.0 MB | 3e53422a50d3 |
| training/checkpoint-4360/rng_state.pth | Weights | 14.6 KB | c72427f97e81 |
| training/checkpoint-4360/scheduler.pt | Weights | 1.5 KB | 54dc364ea571 |
| training/checkpoint-4360/training_args.bin | Weights | 5.8 KB | f102db021651 |
| training/checkpoint-6540/adapter_model.safetensors | Weights | 223.4 MB | acb5d01c2f9a |
| training/checkpoint-6540/optimizer.pt | Weights | 447.0 MB | d8b0a7c52f96 |
| training/checkpoint-6540/rng_state.pth | Weights | 14.6 KB | 7ecff8432e52 |
| training/checkpoint-6540/scheduler.pt | Weights | 1.5 KB | 585326863b1a |
| training/checkpoint-6540/training_args.bin | Weights | 5.8 KB | f102db021651 |
| training/checkpoint-8718/adapter_model.safetensors | Weights | 223.4 MB | 7e404dcac763 |
| training/checkpoint-8718/optimizer.pt | Weights | 447.0 MB | 61b92b1f2c11 |
| training/checkpoint-8718/rng_state.pth | Weights | 14.6 KB | ca0af8fb33f5 |
| training/checkpoint-8718/scheduler.pt | Weights | 1.5 KB | bc8efb8ff5c3 |
| training/checkpoint-8718/training_args.bin | Weights | 5.8 KB | f102db021651 |
| training/checkpoint-best/adapter_model.safetensors | Weights | 223.4 MB | acb5d01c2f9a |
| training/checkpoint-best/optimizer.pt | Weights | 447.0 MB | d8b0a7c52f96 |
| training/checkpoint-best/rng_state.pth | Weights | 14.6 KB | 7ecff8432e52 |
| training/checkpoint-best/scheduler.pt | Weights | 1.5 KB | 585326863b1a |
| training/checkpoint-best/training_args.bin | Weights | 5.8 KB | f102db021651 |
| adapter/adapter_config.json | Configuration | 11.2 KB | — |
| adapter/data_quality.json | Configuration | 1.0 KB | — |
| adapter/kuza_system_prompt.json | Configuration | 883 B | — |
| adapter/sft_manifest.json | Configuration | 10.7 KB | — |
| adapter/train_metrics.json | Configuration | 198 B | — |
| adapter/trainer_state.json | Configuration | 8.8 KB | — |
| imatrix/corpus_manifest.json | Configuration | 495 B | — |
| merged_bf16/config.json | Configuration | 3.0 KB | — |
| merged_bf16/generation_config.json | Configuration | 141 B | — |
| merged_bf16/kuza_system_prompt.json | Configuration | 883 B | — |
| merged_bf16/model.safetensors.index.json | Configuration | 54.8 KB | — |
| provenance/adapter_config.json | Configuration | 11.2 KB | — |
| provenance/data_quality.json | Configuration | 1.0 KB | — |
| provenance/kuza_system_prompt.json | Configuration | 883 B | — |
| provenance/provenance_index.json | Configuration | 177 B | — |
| provenance/q4_k_m_imatrix_recipe.json | Configuration | 672 B | — |
| provenance/q4_k_s_ssm_recipe.json | Configuration | 683 B | — |
| provenance/q4_k_xl_ssm_recipe.json | Configuration | 706 B | — |
| provenance/sft_manifest.json | Configuration | 10.7 KB | — |
| provenance/train_metrics.json | Configuration | 198 B | — |
| provenance/trainer_state.json | Configuration | 8.8 KB | — |
| quants/q4_k_m_imatrix/recipe.json | Configuration | 672 B | — |
| quants/q4_k_s_ssm/recipe.json | Configuration | 683 B | — |
| quants/q4_k_xl_ssm/recipe.json | Configuration | 706 B | — |
| reference/kuza_system_prompt.json | Configuration | 883 B | — |
| reference/reference_manifest.json | Configuration | 668 B | — |
| screen/mtp_probe.json | Configuration | 324 B | — |
| training/checkpoint-2180/adapter_config.json | Configuration | 11.2 KB | — |
| training/checkpoint-2180/trainer_state.json | Configuration | 3.0 KB | — |
| training/checkpoint-4360/adapter_config.json | Configuration | 11.2 KB | — |
| training/checkpoint-4360/trainer_state.json | Configuration | 4.9 KB | — |
| training/checkpoint-6540/adapter_config.json | Configuration | 11.2 KB | — |
| training/checkpoint-6540/trainer_state.json | Configuration | 6.8 KB | — |
| training/checkpoint-8718/adapter_config.json | Configuration | 11.2 KB | — |
| training/checkpoint-8718/trainer_state.json | Configuration | 8.6 KB | — |
| training/checkpoint-best/adapter_config.json | Configuration | 11.2 KB | — |
| training/checkpoint-best/trainer_state.json | Configuration | 6.8 KB | — |
| upload_manifest.json | Configuration | 20.4 KB | — |
| README.md | Documentation | 2.1 KB | — |
| adapter/README.md | Documentation | 5.2 KB | — |
| training/README.md | Documentation | 1.4 KB | — |
| training/checkpoint-2180/README.md | Documentation | 5.2 KB | — |
| training/checkpoint-4360/README.md | Documentation | 5.2 KB | — |
| training/checkpoint-6540/README.md | Documentation | 5.2 KB | — |
| training/checkpoint-8718/README.md | Documentation | 5.2 KB | — |
| training/checkpoint-best/README.md | Documentation | 5.2 KB | — |
| adapter/chat_template.jinja | Other | 9.0 KB | — |
| adapter/heldout.jsonl | Other | 6.5 MB | — |
| imatrix/calibration.txt | Other | 1.9 MB | — |
| imatrix/eval.txt | Other | 397.0 KB | — |
| imatrix/imatrix.log | Other | 1.2 KB | — |
| imatrix/kuza.imatrix | Other | 3.6 MB | 87adb974a66d |
| merged_bf16/chat_template.jinja | Other | 9.0 KB | — |
| quants/q4_k_m_imatrix-smoke.log | Other | — | |
| quants/q4_k_m_imatrix.log | Other | 67.9 KB | — |
| quants/q4_k_s_ssm-smoke.log | Other | — | |
| quants/q4_k_s_ssm.log | Other | 67.9 KB | — |
| quants/q4_k_xl_ssm-smoke.log | Other | — | |
| quants/q4_k_xl_ssm.log | Other | 80.8 KB | — |
| reference/smoke.log | Other | — | |
| screen/bf16-kld.log | Other | 999 B | — |
| screen/kuza-bf16-reference.kld | Other | 4.1 GB | 901bc84b9f91 |
| screen/q4_k_m_imatrix-bench.log | Other | 790 B | — |
| screen/q4_k_m_imatrix-english_agriculture.log | Other | — | |
| screen/q4_k_m_imatrix-english_safety.log | Other | — | |
| screen/q4_k_m_imatrix-kld.log | Other | 3.9 KB | — |
| screen/q4_k_m_imatrix-swahili_agriculture.log | Other | — | |
| screen/q4_k_m_imatrix-swahili_safety.log | Other | — | |
| training/checkpoint-2180/chat_template.jinja | Other | 9.0 KB | — |
| training/checkpoint-4360/chat_template.jinja | Other | 9.0 KB | — |
| training/checkpoint-6540/chat_template.jinja | Other | 9.0 KB | — |
| training/checkpoint-8718/chat_template.jinja | Other | 9.0 KB | — |
| training/checkpoint-best/chat_template.jinja | Other | 9.0 KB | — |
| .gitattributes | Repository | 2.4 KB | — |
| adapter/tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc |
| adapter/tokenizer_config.json | Tokenizer | 7.2 KB | — |
| merged_bf16/tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc |
| merged_bf16/tokenizer_config.json | Tokenizer | 1.2 KB | — |
| training/checkpoint-2180/tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc |
| training/checkpoint-2180/tokenizer_config.json | Tokenizer | 7.2 KB | — |
| training/checkpoint-4360/tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc |
| training/checkpoint-4360/tokenizer_config.json | Tokenizer | 7.2 KB | — |
| training/checkpoint-6540/tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc |
| training/checkpoint-6540/tokenizer_config.json | Tokenizer | 7.2 KB | — |
| training/checkpoint-8718/tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc |
| training/checkpoint-8718/tokenizer_config.json | Tokenizer | 7.2 KB | — |
| training/checkpoint-best/tokenizer.json | Tokenizer | 20.0 MB | 87a7830d63fc |
| training/checkpoint-best/tokenizer_config.json | Tokenizer | 7.2 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 29.0 GB
Released by Kuza AI through its official repository on Hugging Face. Read the license.
Built From
- Adapter of unsloth/Qwen3.5-4B
- Derived from unsloth/Qwen3.5-4B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 29.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About kuza-qwen-3.5-4b
Can I use kuza-qwen-3.5-4b commercially?
Yes. kuza-qwen-3.5-4b is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.