Alpha
Contact
Use · MORENA 1.5B

Is it good enough?

Ownership levelPartialnone·limited·partial·substantial·fullAnalytical input D ยท 46.8/100

This page is a projection of the one entry record, the Reliability factor that Use covers. The full verdict is set by all four factors together, floor-weighted so the weakest caps the whole.

Which domain expands which factor
  • AssessUse & modify + Transparency
  • ImplementData control + Reliability
  • UseReliability
  • SupportTransparency

Capabilities & modalities

Text only. The measured strengths are African-language modelling (lowest bits per byte of 26 models in the paper) and English-to-African translation (chrF++ 45.8 across five languages, 3-shot). Code and French come from the mixture.

The weak spots are comprehension (Belebele 0.309, retrieval QA 0.325 against a 0.25 floor), grounded faithfulness (23%), English knowledge (MMLU 0.284, deliberately traded away) and over-refusal.

Its tokenizer suits African text: the paper reports 1.39 times fewer tokens than under Gemma 3's vocabulary and 1.53 times fewer than under Llama 3.2's (ev-paper). All figures are the publisher's own.

Context window & long-context behaviour

The context window is 4,096 tokens, as stated on the card and the paper and in config.json (seq_len 4096), tokenizer_config.json (model_max_length 4096) and the GGUF header read by Hugging Face (ev-card-instruct, ev-hf-platform). No long-context training or evaluation is described. Chunk or retrieve instead of stuffing documents, and note that retrieval QA is itself near chance.

Prompt format & chat template

Turns are marked by single reserved tokens, <reserved_0> for the user (id 3) and <reserved_1> for the assistant (id 4). tokenizer_config.json (added 2026-10-14) holds a Jinja template. It renders a BOS of <eos>, an optional system message followed by a newline, then <reserved_0>, a newline, the user text and a newline. The generation prompt is <reserved_1> and a newline. An assistant turn ends with <eos>.

The card warns that <|user|>-style strings "produce degenerate output", and the paper says that mistake once invalidated an internal safety evaluation (ev-card-instruct, ev-trainer). The base checkpoint has no chat format and should not be prompted as an assistant.

Language coverage

The cards name twelve African languages, plus English, French and code (ev-card-instruct, ev-paper). The evidence differs by language, so the list does not imply equal support.

LanguageNative-text or machine-translated pretrainingHeld-out evaluation textTranslation scoredSafety weakness
SwahiliMT poolFLORES+, MasakhaNEWSyes (59.9 into)no
HausaMT poolFLORES+, MasakhaNEWSyes (49.1 into)no
Yorubano MT pool in Table 2FLORES+, MasakhaNEWSyes (26.9 into)82%
IgboMT poolFLORES+, MasakhaNEWSno75%
isiZuluMT poolFLORES+, NCHLTyes (48.0 into)no
isiXhosaMT poolFLORES+, MasakhaNEWS, NCHLTyes (45.1 into)87%
ShonaMT poolFLORES+, MasakhaNEWSnono
KinyarwandaMT poolFLORES+nono
SetswanaMT poolFLORES+, NCHLTno83%
AfrikaansMT poolFLORES+, NCHLTnono
isiNdebelenone, plus under 40M synthetic tokensNCHLT transcripts onlynonot listed
Nigerian Pidginnone, plus under 40M synthetic tokensMasakhaNEWS onlynonot listed

Translation scores are chrF++ from English from the paper's per-language table. Safety figures are the weakest-four list in the paper. The paper's text says only isiNdebele and Nigerian Pidgin lack a machine-translated pool, but Table 2 lists none for Yoruba either. Press articles repeat the publisher's language list and add no evidence.

Function / tool calling

The instruct model reaches 98.1% correct tool and 100% valid JSON. The paper says this is "format compliance and tool selection given the decision to call a tool", because the tool marker is prefilled in the test. Left to decide on its own, the model almost never calls a tool (ev-card-instruct, ev-paper). Tool use is unproven unless the caller prefills or constrains the call.

How this scores

The ownership factor this domain covers, drawn from the one entry record.

3

ReliabilityIs it reliable and good enough for the job?

Weak

Operational 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).

How this scores (AOI sub-dimensions)
Operational2/5how practical it is to run, serve and maintain in productionThe only mainstream route is the publisher's GGUF on llama.cpp (Q4_K_M, Q8_0, f16; Hugging Face reads it as architecture llama with 4,096 context).
Safety2/5whether misuse risks are evaluated and guardrails are providedThe instruct model is safety-tuned, and the publisher reports a 3,341-prompt, 13-language, 11-category evaluation that includes its residual risks (89.0% handled well, 41% of benign requests failed, Igbo 75%).
What this means for adoptionYou own the use of the weights: Apache-2.0 is declared, ungated and commercial use is stated, and your data stays on your hardware. You do not own a finished assistant. The paper says the instruct model is not one, it cannot be rebuilt without the unreleased corpus and configs, and the evaluations are the publisher's own. The level rises if Vambo ships a LICENSE file, releases its corpus and harness, and clears the Gemma Terms question on the SFT data.

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.

Model cardread2026-11-19
The instruct card declares Apache-2.0, names twelve languages, and gives 1.485B parameters and a 4,096-token context.
Model cardread2026-11-19
The base card describes a base checkpoint beside the instruct one (vamboai/morena-1.5b-base) with the same Apache-2.0 declaration.
Licenceread2026-11-19
Every MORENA repo declares Apache-2.0 in the front matter and the card 'Licence' section ('Commercial use is permitted, including fine-tuning and redistributing derivatives').
Licenceread2026-11-19
Canonical Apache License 2.0 text, the licence the MORENA declaration points to: section 2 grants copyright, section 3 grants patent rights, and section 4 governs redistribution and requires passing on a copy of the licence.
Technical_reportread2026-11-19
Paper 'MORENA: An African Foundation Model' (23 pages, 18 September 2026, single author).
Documentationread2026-11-19
modeling_morena.py (52,103 bytes, identical in base and instruct) calls itself the 'Morena production trainer' (FSDP2, Muon, mixture sampler, checkpointing) and also defines the model.
Documentationread2026-11-19
SHA256SUMS in the repo lists five files.
Documentationread2026-11-19
Hugging Face platform metadata.
Model cardread2026-11-19
Sibling cards: morena-0.5b-mini and morena-0.2b-nano (pruned and distilled from the 1.5B base, not safety-tested, Apache-2.0) and morena-1.5b-instruct-gguf (Q4_K_M 0.92 GB, Q8_0 1.58 GB, f16 2.97 GB, 'verified for logit equivalence').
Documentationread2026-11-19
vamboai/fikira is a public, ungated MIT dataset of 50,000 synthetic reasoning examples in ten languages, copyright 'Vambo Technologies (trading as Vambo AI)', one named SFT component; vamboai/morena-sft-corpus, listed in the card metadata, returns HTTP 401 and is not public.
Vendor announcementread2026-11-19
The vambo.ai footer names Vambo Technologies (Pty) Ltd, and the site names co-founders.
Third-party analysisread2026-11-19
Three press articles (Tech In Africa 13 Oct 2026, CIO Africa 16 Oct 2026, Daba Finance 7 Oct 2026) describe a South African start-up releasing an open-weight model and say the results are the developer's own and unverified.
Terms of serviceread2026-11-19
Gemma Terms of Use (last modified April 1, 2026).