An offline enterprise assistant for the laptops West Africa actually owns: 8 GB of RAM, integrated graphics, no network. It reads a company's own documents — supplier contracts, invoices, HR policies, meeting notes — and answers with citations, on the machine, in French and English.

What this file is, exactly

This repository hosts Qwen3.5-2B-IQ4_XS.gguf, an unmodified copy of the IQ4_XS build published by unsloth/Qwen3.5-2B-GGUF, itself quantised from Qwen/Qwen3.5-2B. Credit for the base model goes to Qwen; credit for this quantisation goes to Unsloth. It is republished here so that the submission's download_model.sh points at a URL under our control and keeps working unchanged through the audit window — not because we claim authorship of the weights.

sha256  3639f34b5ca22aa1c51f3616566eae8c355111554f6924ad97ee2652ed11c1cd
size    1 172 996 352 bytes (1.09 GiB)

Our contribution is the selection, measurement and packaging: which base model, which quantisation, which sampling settings, and the evidence for each. That evidence lives in the GitHub repository, not in a claim on this page.

We did try to become authors of these weights rather than their measurers, by recalibrating the importance matrix on francophone enterprise text instead of the inherited generic-English set. It made no measurable difference: on 200 arc_easy questions the two files return 196 identical answers, at identical size, speed and peak memory. So these weights are still Unsloth's, and we say so. The protocol — calibration corpus, contamination check, control file and paired test — is in bench/resultats.md, step 6.

Why this model, and why this quantisation

Both decisions were measured with the official adtc-profiler, never chosen by reputation. Five base models, from 0.75 B to 4.21 B measured parameters, were profiled; then all seven quantisations of the winner. Full tables: bench/resultats.md.

Measured on an Apple M1 / 8 GB, CPU only (-ngl 0, enforced by the profiler). Throughput and peak memory are the median of three runs — a single memory reading is worthless, and we have the scar to prove it: one variant showed 1.47 GB on its first pass and 2.21 GB as its true median. Accuracy is a single deterministic run (temperature 0, fixed seed, 200 arc_easy questions); repeating it would return the same number. Absolute values differ from the reference i5 laptop; the ranking between candidates does not.

Quantisation	Accuracy	Throughput	Peak RAM	S_eff	Total @150 t/s
IQ4_XS (shipped)	0.670	34.3 t/s	1.74 GB	75.2	55.4
Q4_K_M	0.675	31.6 t/s	2.08 GB	70.2	54.1
UD-Q5_K_XL	0.680	29.0 t/s	2.11 GB	69.8	53.8
MTP-Q4_K_M	0.675	31.1 t/s	2.20 GB	68.5	53.7
Q5_K_M	0.670	26.7 t/s	2.01 GB	71.3	53.1
UD-Q4_K_XL	0.650	29.4 t/s	2.21 GB	68.4	52.1
Q3_K_M	0.630	30.7 t/s	1.93 GB	72.4	52.1

The last column is not a measurement: it is the official scoring function 0.50·accuracy + 0.30·S_perf + 0.20·S_eff applied to the measured cells, under the assumption that the fastest submission in the contest reaches 150 t/s. S_perf is scored relative to that submission, so the assumption has to be stated rather than hidden.

The variant that beats us is in the table on purpose. UD-Q5_K_XL scores 0.680 against our 0.670 — the best accuracy of the seven. It still loses overall, and the arithmetic says by how much: that extra point of accuracy is worth 0.5 of final score, while the 18 % throughput and 5.4 S_eff it gives up cost it 2.1 — a net 1.6 in our favour, which is exactly the 55.4 against 53.8 in the table. That is the whole argument for this track in one row, and hiding the row would have made the argument weaker, not stronger.

IQ4_XS is also the fastest and the lightest, and its three runs sat within 1.72–1.77 GB — the narrowest spread we recorded, which matters because it is the figure that has to survive an independent re-measurement.

The shipped file, measured as a package

The table above ranks candidates. The number that describes this file as it is submitted — fetched by download_model.sh, three profiler runs, median — is 31.20 t/s and 1 544 MB peak. It is lower than the 34.3 t/s above and that is not a contradiction to explain away: it is run-to-run and thermal variance on a fanless 8 GB laptop, measured weeks apart. The ranking table is used only to compare variants measured back to back; the packaged figure is the one we self-report.

Recommended inference settings

llama-cli -m Qwen3.5-2B-IQ4_XS.gguf -ngl 0 --temp 0 --repeat-penalty 1.05

--repeat-penalty 1.05 is not a preference. On inputs outside its competence this model does not decline — it repeats one phrase until the token budget runs out, and llama.cpp applies no repetition penalty by default.

The value was measured twice, and the second measurement overturned the first. An arithmetic control of 18 items pointed at 1.10. A second control of 15 summarisation, drafting and analysis tasks — the genre this model is actually for — showed what that had cost. On a contract-penalty question, 1.00 and 1.05 both produce 270,000 FCFA, a defensible amount; 1.10 produces 63,450 FCFA by inventing a formula, (30 − 25) / 7, that corresponds to nothing in the contract. Reproducible at temperature 0.

To be precise about what "defensible" means here, because it is not the same as right: 270,000 follows if the ten-day threshold is read as a grace period, leaving 15 days — three weeks begun — at 2 % each. The model does not reason that way. It divides 25 by 7, gets "3 weeks and 4 days", and calls that three weeks begun, which rounds a begun week down. It reaches a defensible number by an indefensible route. That rounding failure is listed under limitations below and it is not fixed by any penalty value.

1.05 keeps the model on that route rather than the fabricated one, still removes the degeneration (diversity 0.60 → 0.99 on the case that showed it), and costs one criterion out of 81 against no penalty at all. Above 1.10 the collapse is not subtle: multi-step reasoning falls from 9/12 to 4/12 at 1.15.

Sweeps and full transcripts: bench/copies/redaction.md, bench/copies/penalite-repetition.md.

Known limitations, measured

No African-language capability. Probed and documented: asked to identify Dioula it answered "the language of Cameroon"; asked for Wolof, "the language of Tigré". dyu was consequently removed from the submission's language_scope. Working languages are French and English.
Rounding to a week begun — a common clause in West African supply contracts — is wrong at every configuration we tested. It rounds down: 25 days becomes "three weeks begun".
It drops a fact to make room for a comment. Told to summarise a clinic report in exactly three bullets, it sacrificed the 71 % bed-occupancy figure to write "requires immediate intervention". Summarising a client thread, it never quoted the order reference.
It ranks urgency badly. Asked to order four tasks, it placed a public tender closing in three days last, as "low urgency" — it had restated the order of the question with justifications attached.
It confuses accounting definitions, computing gross margin as revenue minus fixed costs.
It can derive numbers confidently and wrongly. Analysing a purchasing proposal, it divided an annual spend by 1.08 to "recover" a pre-saving baseline, then built two further figures on that false start.

None of these depend on sampling settings; they are in the base model. The last four were found by bench/redaction.py, a 15-task control scored without human judgement.

2 B parameters is a deliberate trade, not a limitation we are apologising for. Half of the score is throughput and memory. Measured on the same machine at the Q4_K_M stage, Qwen3.5-4B is 6 accuracy points better — 0.735 against 0.675 — and still loses on total score, 52.6 against 56.5 in the same 150 t/s scenario, because it runs at 44 % of the speed and takes 1.4× the memory.

Licence

Apache 2.0, inherited from Qwen3.5-2B. The submission repository is GPL v3, inherited from the official ADTC template; the weights keep their own licence.

Benewende-dev/baarali-edge-2b

baarali-edge-2b