Version: v0.1.23-alpha (project framework) Last Updated: 2026-05-25 Developer: Smouj (@smouj013)
Planned / Training Design
Kimari-4B has not been trained yet. No model weights exist, no training run has been started, and no benchmarks have been measured. This model card describes the design targets and planned architecture for a future fine-tuned local LLM. All specifications below are aspirational and subject to change during development.
Important: Until Kimari-4B is released, the Kimari framework can run compatible GGUF base models for testing. See
models/README.mdfor recommended test models.
No weights have been released. Kimari-4B model weights do not exist in any form — no safetensors, no GGUF, no Hugging Face repository, no download link. Any claim of available Kimari-4B weights is false.
When weights are eventually produced, they will be announced through official channels and this model card will be updated accordingly.
| Field | Value |
|---|---|
| Model Name | Kimari-4B |
| Model Type | Causal Language Model (Transformer) — fine-tuned for local coding, system administration, and agent assistance |
| Target Parameter Class | 3B–4B parameters |
| Target Context Length | Up to 16,384 tokens (profile-dependent) |
| Intended Runtime | llama.cpp / llama-server via GGUF quantization |
| Primary Languages | English, Technical Spanish |
Kimari-4B is designed to run entirely on consumer-grade GPUs available since 2016. The following table defines the minimum and recommended hardware targets:
| GPU | VRAM | Recommended Quantization | Target Context | Notes |
|---|---|---|---|---|
| NVIDIA GTX 1060 | 6 GB | GGUF Q4_K_M | 8,192 tokens | Minimum viable GPU; safe VRAM budget ~4.9 GB |
| NVIDIA GTX 1080 | 8 GB | GGUF Q5_K_M | 16,384 tokens | Comfortable Q5 fit; safe VRAM budget ~6.8 GB |
| NVIDIA RTX 2060 | 6 GB | GGUF Q4_K_M | 8,192 tokens | Similar to GTX 1060 with newer architecture |
| NVIDIA RTX 3060 | 12 GB | GGUF Q5_K_M | 32,768 tokens | Extended context; room for larger batches |
These are design targets, not guaranteed performance levels. Actual performance will depend on the final base model, quantization behavior, and training outcome.
| Format | Status | Notes |
|---|---|---|
| safetensors (full base or adapter) | Planned | Availability depends on selected base model and license |
| GGUF Q4_K_M | Planned | Primary distribution format for 6 GB VRAM targets |
| GGUF Q5_K_M | Planned | Higher quality for 8 GB+ VRAM targets |
| GGUF IQ4_XS | Optional | May be produced if quality degradation is acceptable for maximum compression |
Format availability is contingent on the selected base model's license terms. Some licenses may restrict distribution of derivative weights or require specific attribution.
The base model has not been selected yet. The following candidates are under evaluation. Each has distinct trade-offs in capability, licensing, and community ecosystem:
| Candidate | Parameters | License | Strengths | Concerns |
|---|---|---|---|---|
| SmolLM3-3B | ~3B | Apache 2.0 | Fully permissive license; allows unrestricted distribution of fine-tuned weights and commercial use; active community | Smaller parameter count may limit reasoning depth; relatively newer model with less community benchmarking |
| Qwen2.5-3B-Instruct | ~3B | Requires review | Strong multilingual and coding capabilities; well-benchmarked; instruction-tuned out of the box | License must be reviewed before commitment — may impose restrictions on derivative distribution or commercial use |
| Llama 3.2 3B | ~3B | Llama 3.2 Community License | Strong general capability; Meta's training infrastructure; large ecosystem | Custom Meta license with specific use restrictions (e.g., monthly active user thresholds, acceptable use policy); may constrain distribution of fine-tuned weights |
The final base model selection will be based on:
- License compatibility — Must permit distribution of fine-tuned weights in planned formats (GGUF, safetensors)
- Capability baseline — Pre-training quality in coding, reasoning, and multilingual understanding
- Community benchmarks — Available evaluation data (MMLU, HumanEval, etc.) for the base model
- Tokenization alignment — No mixing of models/tokenizers across different architectures
- Ecosystem support — llama.cpp / GGUF conversion maturity for the architecture
No base model has been definitively chosen. The "Selected Base" field below will be updated when a final decision is made.
HuggingFaceTB/SmolLM3-3B — Accepted for first private SFT training run. NOT yet accepted for public release. Public release base subject to evaluation results and full license review. See docs/BASE_MODEL_ACCEPTANCE.md for details.
TBD — The license for Kimari-4B weights depends entirely on the selected base model:
- If SmolLM3-3B (Apache 2.0) is selected → Kimari-4B weights could be released under Apache 2.0
- If Qwen2.5-3B-Instruct is selected → License TBD pending review of Qwen's terms
- If Llama 3.2 3B is selected → Llama 3.2 Community License applies, with its specific restrictions on derivative distribution and commercial use
The Kimari software (CLI, scripts, configurations) is released under the MIT License regardless of the model license. See MODEL_LICENSES.md for full details on license layers.
Do not assume any specific license for Kimari-4B weights until the base model is selected and license compatibility is confirmed.
| Aspect | Status |
|---|---|
| Base model selection | Accepted for private SFT (SmolLM3-3B) |
| Training data curation | Not started |
| Fine-tuning run | Not started |
| Evaluation | Not started |
| Weight release | Not started |
Training has not begun. No GPUs have been allocated, no training scripts have been executed, and no checkpoints exist. The project is currently in the training design phase — defining requirements, evaluating base models, and planning data curation.
| Aspect | Status |
|---|---|
| Current model work | v0.1.21 adapter manifest template, eval summary policy, SFT→ORPO decision |
| Base selection status | SmolLM3-3B accepted for first private SFT candidate |
| Dataset status | v0 synthetic dataset prepared (SFT + preference + holdout) |
| Training status | Not started (configs and validation ready) |
| Evaluation status | KimariFit scoring plan prepared, dimensions defined |
| GGUF export | Plan defined (Q4_K_M, Q5_K_M, IQ4_XS) |
| HF release | Placeholder plan defined, not executed |
No benchmarks have been measured. Kimari-4B does not exist as a trained model, so no benchmark scores are available.
The following table lists evaluation targets only — these are aspirational goals for the future model, not achieved results:
| Benchmark | Metric | Q4_K_M Target | Q5_K_M Target | Notes |
|---|---|---|---|---|
| Inference Speed (GTX 1060) | tokens/second | > 15 t/s | N/A (6 GB limited) | Target for Q4_K_M on minimum hardware |
| Inference Speed (GTX 1080) | tokens/second | > 20 t/s | > 15 t/s | Q5 fits comfortably on 8 GB |
| Time to First Token (TTFT) | seconds | < 2s | < 3s | Prompt processing latency |
| MMLU (5-shot) | accuracy | TBD | TBD | Target to be set after base model selection |
| HumanEval (pass@1) | accuracy | TBD | TBD | Core coding benchmark target |
| MultiPL-E (Python) | pass@1 | TBD | TBD | Coding capability target |
| SWE-bench (Lite) | resolved % | TBD | TBD | Agent/reasoning capability target |
These are targets, not results. Actual performance may differ significantly. Benchmark targets will be refined after base model selection and initial evaluation of the base model's capabilities.
- Local coding assistant — Code generation, debugging, code review, and explanation in Python, TypeScript, shell scripting, and other languages
- System administration — Linux and Windows troubleshooting, automation script generation, log analysis, and infrastructure guidance
- Agent orchestration — Multi-step reasoning for semi-autonomous tasks, tool use planning, and decision chains
- Technical documentation — Writing and explaining technical concepts, generating documentation, and summarizing complex topics
- Bilingual assistance — English and technical Spanish support for IT infrastructure, DevOps, and software engineering
- Developers working locally without internet access (air-gapped environments)
- System administrators needing quick, private troubleshooting assistance
- Teams with data privacy requirements requiring on-premise AI
- Students and learners in technical fields
- Users with consumer-grade GPUs (GTX 1060 or better)
The following uses are explicitly out of scope for Kimari-4B:
- Medical, legal, or financial advice — The model is not qualified for professional advice in regulated domains
- Real-time safety-critical systems — Do not use for medical diagnosis, autonomous vehicle control, or any system where failure could cause harm
- Surveillance or privacy violation — Do not use to process personal data without consent or for surveillance purposes
- Generation of harmful content — Malware, exploits, harassment, disinformation, or any content intended to cause harm
- Replacing human expertise in critical decisions — Model outputs must always be reviewed by a qualified human
- Large-scale commercial deployment without license review — Ensure base model license permits your intended use case
- Multimodal tasks — Kimari-4B is text-only; it does not process images, audio, or video
- General-purpose knowledge tasks beyond technical domains — The model is specialized for coding/sysadmin/agent tasks; general knowledge accuracy is not a design priority
- No real-time knowledge — The model has no access to the internet, APIs, or current events. All knowledge is frozen at training time.
- Hallucination risk — Like all LLMs, Kimari-4B may generate plausible but factually incorrect information, especially in domains outside its training focus.
- Context window constraints — Available context length is limited by VRAM. Long documents or conversations may be truncated. On 6 GB GPUs, context is limited to ~8K tokens.
- Quantization artifacts — Lower quantization levels (Q4_K_M, IQ4_XS) may degrade output quality, especially for nuanced reasoning or precise code generation.
- No multimodal capability — Text-only input and output. Cannot process images, diagrams, audio, or other modalities.
- Language coverage — Optimized for English and technical Spanish. Other languages are not tested or supported.
- Reasoning depth — As a 3B–4B parameter model, complex multi-step reasoning may be inconsistent or incorrect. Larger models will generally outperform on difficult reasoning tasks.
- Code correctness — Generated code may contain bugs, security vulnerabilities, or deprecated APIs. Always review and test generated code before use.
- Base model limitations inherited — Whatever limitations the selected base model has (biases, knowledge gaps, safety concerns) will be inherited by Kimari-4B.
- No self-correction guarantee — The model cannot reliably identify its own mistakes. External validation is always required.
- Model outputs should never be used for critical decisions without human review
- Generated code must be reviewed for security vulnerabilities before deployment
- No PII is collected or transmitted — all processing is local by design
- Model weights are not included in this repository
| Risk | Description | Mitigation |
|---|---|---|
| Misinformation | Model may generate factually incorrect content with high confidence | Always verify outputs against authoritative sources |
| Bias | Training data may contain biases that propagate to outputs | Be aware of potential bias; do not use for decisions affecting people |
| Code vulnerabilities | Generated code may contain security flaws | Mandatory human code review before any deployment |
| Over-reliance | Users may trust model outputs without verification | Treat model as a suggestion tool, not an authority |
| Prompt injection | Malicious prompts may manipulate outputs in unexpected ways | Sanitize inputs in production; do not pipe untrusted input directly |
Users are responsible for:
- Reviewing all model outputs before use in any context
- Not using the model for harmful, illegal, or unethical purposes
- Complying with applicable laws and regulations in their jurisdiction
- Ensuring base model license compliance for their specific use case
- Implementing appropriate safeguards when deploying in shared environments
The following items must be completed before Kimari-4B weights can be released:
| # | Item | Status |
|---|---|---|
| 1 | Base model selected and license reviewed | 🟡 In Progress (SmolLM3-3B accepted for private training) |
| 2 | License compatibility confirmed for all planned formats | ❌ Not started |
| 3 | Seed dataset v0 prepared and documented | 🔶 In Progress |
| 3b | Full training dataset curated | ❌ Not started |
| 4 | Fine-tuning completed | ❌ Not started |
| 5 | Base model benchmarks measured (pre-fine-tune baseline) | ❌ Not started |
| 6 | Fine-tuned model benchmarks measured | ❌ Not started |
| 7 | GGUF quantization produced and validated (Q4_K_M, Q5_K_M) | ❌ Not started |
| 8 | Optional IQ4_XS quantization evaluated | ❌ Not started |
| 9 | Inference speed tested on all target GPUs | ❌ Not started |
| 10 | Safety evaluation completed | ❌ Not started |
| 11 | Model card updated with actual benchmark results | ❌ Not started |
| 12 | MODEL_LICENSES.md updated with confirmed license | ❌ Not started |
| 13 | Hugging Face repository created (or equivalent distribution) | ❌ Not started |
| 14 | Kimari framework model registry updated | ❌ Not started |
| 15 | Community announcement published | ❌ Not started |
This checklist will be updated as progress is made. No item has been started as of v0.1.17-alpha.
As of v0.1.19-alpha, SmolLM3-3B has been accepted for first private SFT training. No real training, evaluation, or public release has occurred.
| Version | Date | Status | Changes |
|---|---|---|---|
| 0.1.23-alpha | 2026-05-25 | Current | Postrun --json fix; preflight config-aware dataset_build_dir; screenshots documentation; release-check improvements |
| 0.1.22-alpha | 2026-05-24 | Released | Private SFT execution package; remote GPU guide; preflight/postrun scripts; training requirements; private run artifacts/failures policy; baseline/adapter eval plan scripts |
| 0.1.21-alpha | 2026-05-23 | Released | Adapter manifest template; eval summary policy; SFT→ORPO decision framework; private SFT execution checklist; create_adapter_manifest and create_eval_summary scripts |
| 0.1.20-alpha | 2026-05-22 | Released | Private SFT runbook; baseline eval plan; adapter artifact policy; preview gate; compare runs tool; pipeline dry-run orchestration |
| 0.1.19-alpha | 2026-05-21 | Released | SmolLM3-3B accepted for first private SFT candidate; dataset v0; training readiness validation; KimariFit scoring plan; v0 training configs; HF placeholder plan |
| 0.1.18-alpha | 2026-05-20 | Released | Pipeline dry-run: base decision record, seed datasets, dataset mix builder, KimariFit dry-run harness, GGUF export plan |
| 0.1.17-alpha | 2026-05-19 | Released | Complete model card rewrite for v0.1.17-alpha; transparent status, honest evaluation targets, expanded safety and limitation sections |
| — | TBD | Future | Initial training design document based on selected base model |
| — | TBD | Future | First fine-tuning run and benchmark results |
| — | TBD | Future | Weight release (safetensors + GGUF) |
This model card will be updated as development progresses. No information herein should be interpreted as a promise or guarantee of future results.