Can you run it?
This page is a projection of the one entry record, the Reliability and Data control factors that Implement covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Install & run
The weights are not a transformers architecture, so AutoModelForCausalLM cannot load them. The publisher route is its own PyTorch file (ev-card-instruct, ev-trainer):
pip install torch safetensors tokenizers
# download model.safetensors, config.json, tokenizer.json, modeling_morena.py,
# load_example.py and SHA256SUMS from vamboai/morena-1.5b-instruct at a pinned revision
sha256sum -c SHA256SUMS # confirm all five files before running anything
python load_example.py # needs a CUDA GPU as written
load_example.py builds Transformer(ModelConfig(**cfg["model"]), "sdpa"), loads the weights with strict=True and samples with a hand-written loop. The class has no .generate(). The other route is the official GGUF (vamboai/morena-1.5b-instruct-gguf) with llama.cpp:
llama-completion -m morena-1.5b-instruct-f16.gguf \
-p "<reserved_0>\nNdeipi guta guru reZimbabwe?\n<reserved_1>\n" -e -n 64 --temp 0
Pin the revision hash. The instruct repo was last modified on 2026-10-14. modeling_morena.py is code you will execute, so read it first and run it in a sandbox with no network. load_example.py still opens with "preview weights". The text is stale and does not refer to a different checkpoint (ev-hf-platform).
Hardware & VRAM requirements
The publisher gives file sizes and no requirements table (ev-card-instruct, ev-paper, ev-hf-platform).
| Artefact | Size |
|---|---|
| model.safetensors (bf16, 1,484,900,352 parameters) | 2.97 GB |
| GGUF Q4_K_M / Q8_0 / f16 | 0.92 / 1.58 / 2.97 GB |
The paper reports 4.44 ms per character on one A100, measured on the previous instruct version with the same architecture, and says the GGUF builds run on a laptop CPU offline. The 4,096-token context keeps the KV cache small. The reference loop has no KV cache, so it recomputes the whole sequence per token and suits short tests only. OneHill has not measured memory or speed.
Serving stacks
| Stack | Status |
|---|---|
| llama.cpp | Official GGUF. Hugging Face parses it as architecture llama, context 4096 (ev-hf-platform). The publisher says logit equivalence was checked; OneHill has not repeated the check |
| Publisher PyTorch code | Works, with no generate() and no KV cache |
| transformers, vLLM, TGI, Ollama | No documented support or model class |
| Hosted providers | None. The Hugging Face inference mapping is empty and the cards set inference: false |
Operational readiness is therefore 2. Anything beyond llama.cpp means converting the weights or writing a model class.
Safe-deployment controls & Deployment Ceiling
Controls follow the safe-deployment playbook.
Supply chain: pin the revision, verify SHA256SUMS yourself (including the 2.97 GB weights file, which OneHill has not downloaded), and review modeling_morena.py before executing it.
Model layer: use the instruct checkpoint only for experiments, because the base checkpoint has no safety tuning. Expect refusals and degenerate output (1.3%).
Evaluation: have native speakers rate outputs per language before any wider use. Every publisher safety number is one model judging another, and Igbo (75%) and Yoruba (82%) are the weakest languages.
Input and output layers: add your own guardrails. No guard model, acceptable-use policy or jailbreak test ships.
Licence hygiene: add the Apache-2.0 text to anything you redistribute, and record the open Gemma Terms question in your AI-BOM.
Deployment Ceiling: T1 (internal assistant, human reads every output), conditional. The publisher says the instruct model is not yet an assistant, fails 41% of benign requests and has no native-speaker review, so the evidence does not justify customer-facing T2 use.
Available quantizations
The publisher ships GGUF for the instruct model (Q4_K_M, Q8_0, f16), the 0.5B mini and the 0.2B nano (ev-siblings, ev-hf-platform). No GGUF is listed for the 1.5B base, and no GPTQ, AWQ or MLX build exists. The cards say GGUF conversions were checked for logit equivalence with the safetensors. Quality loss at Q4_K_M in the African languages is unmeasured, so test your language.
Fine-tuning & adaptation
The base card recommends the base checkpoint for "fine-tuning, continued pretraining and domain adaptation" (ev-card-base). The repo's modeling_morena.py is the publisher's trainer: FSDP2 sharding, Muon for hidden weights and AdamW for embeddings, a warmup-stable-decay schedule, document-packed sequences, an --anneal mode and a --loss-mask flag for SFT.
It expects configs/*.json, a mixture JSON and sharded data directories that are not published, and it mentions a TRAIN_README.md that is absent. Write your own configs and data loader inputs. The released SFT component is the Fikira dataset (50,000 synthetic rows, MIT); the SFT corpus itself is not public. The paper's instruct recipe, for example 4,500 steps at 2e-5 on batches of 256 with loss masked to assistant turns, is a starting point.
API / OpenAI-compatible integration
There is no hosted endpoint and no documented OpenAI-compatible server. llama.cpp can serve the GGUF over its own HTTP server, but the publisher does not document that route. The prompt format uses reserved tokens, so wrap requests in the chat template from tokenizer_config.json (see the Use domain). Role-based JSON does not apply.
How this scores
The ownership factors this domain covers, drawn from the one entry record.
ReliabilityIs it reliable and good enough for the job?
WeakOperational readiness is 2 (publisher code or llama.cpp GGUF only) and safety is 2 (publisher-only, model-judged evaluation, 41% of benign requests failed). The paper says the instruct model is not an assistant (ev-card-instruct, ev-paper).
Doesn't extract your dataDoes running it keep your knowledge and data yours?
StrongSelf-hosted weights under a declared Apache-2.0 licence with no clawback clause. An outline read of the loader code found no network calls (ev-licence, ev-trainer).
Sources
The same evidence records as the entry sheet. Read means the text was verified; unverified means it is known to exist but not yet read.