Product 01 · Open weights

fihrIA-ASR0.6B & 1.7B

State-of-the-art open-weight speech recognition for Moroccan Darija. The 1.7B makes fewer meaning errors than Mistral’s Voxtral Mini 4B — 6.4% vs 8.4% — and runs about 7× faster; the 0.6B does it on a laptop CPU.

Released2026Speech recognition

At a glance

6.4%
held-out mean SER · 8.4% for Mistral Voxtral, 11.2% for MoulSot
7×
faster than Voxtral on one H100 — 243× vs 34× real time
6.1 / 9.1%
DarijaCS CER, Prism / Probe · Voxtral publishes 6.7 / 9.7%
6×
real time on a laptop CPU · 721 MB at 4-bit

508-clip held-out set, four independent LLM judges, and Morocco’s external DarijaCS benchmark. Voxtral figures are its published numbers.

The models

One family, two sizes

The 1.7B is the accuracy ceiling; the 0.6B is the one that runs without a GPU.

fihrIA-ASR 1.7B

1.7B parameters · Darija BPE

Coming soon

The most accurate open-weight Darija ASR model we tested. It has the lowest meaning-error rate on our 508-clip held-out set and is ahead on both DarijaCS test sets — while running about 7× faster than Voxtral Mini 4B.

Held-out mean SER
6.4%vs 8.4% for Voxtral
DarijaCS CER
6.1% / 9.1%Prism / Probe
H100 throughput
243×real time, batch 32
Footprint
3.8 GB / 1.6 GBBF16 / Q4

The 1.7B weights, GGUF files and release notes are in preparation. The evaluation below is already final.

fihrIA-ASR 0.6B

0.6B parameters · Darija BPE

Available

Compact enough for a laptop, with no GPU. It beats the 1.7B MoulSot baseline under every judge and transcribes at roughly 6× real time on an ordinary CPU.

Held-out mean SER
7.6%vs 11.2% for MoulSot
DarijaCS CER
7.9% / 9.8%Prism / Probe
Laptop CPU
6×real time; 255× on H100
Footprint
1.4 GB / 721 MBBF16 / Q4

Evaluation

How it compares

The 1.7B has the lowest meaning-error rate of any open-weight system we tested, and it beats Mistral’s published DarijaCS numbers on both test sets.

Held-out benchmark · 508 clips · 4 LLM judges

Head to head with MoulSot and Voxtral Mini 4B

Swipe horizontally to compare all four systems →

MoulSot1.7B · Qwen tokenizerVoxtralMini 4B · Realtime ArabicfihrIA-ASR 1.7B1.7B · Darija tokenizerfihrIA-ASR 0.6B0.6B · Darija tokenizer
Parameters1.7B4B1.7B0.6B
Meaning errors (mean SER)11.2%8.4%6.4%7.6%
Median WER42.3%36.5%30.9%31.6%
Median CER13.0%10.6%10.3%10.7%
H100, × real time147×34×243×255×

fihrIA-ASR 1.7B

1.7×fewer meaning errors than MoulSot, at the same size

1.3×fewer meaning errors than Voxtral

1.65×faster than MoulSot

7.1×faster than Voxtral on one H100

fihrIA-ASR 0.6B

1.5×fewer meaning errors than MoulSot, with 2.8× fewer parameters

1.1×fewer meaning errors than Voxtral

2.0×faster than MoulSot

7.5×faster than Voxtral on one H100

Meaning errors under each judge (SER %, lower is better)

Swipe horizontally for every judge →

JudgefihrIA 1.7BfihrIA 0.6BMoulSotVoxtral
GPT-5.6 Luna5.86.910.27.5
Gemini 3.8 Flash7.18.211.99.8
Claude Sonnet 57.08.013.49.7
GLM-5.35.87.29.26.5

MoulSot is the other Qwen3-ASR 1.7B fine-tune for Darija. Voxtral Mini 4B Realtime Arabic is a separate open-weight model. All four were tested on the same 508 clips. Meaning errors and the multiples are the mean of the four judges' SER. The strip below shows each judge, and they do not all rank Voxtral the same way. H100 speed is batch size 32. MoulSot's H100 figure is the control with MoulSot's architecture and tokenizer, and that is also the "faster than MoulSot" ratio for our 1.7B model. For our 0.6B model, "faster than MoulSot" is the laptop comparison, both models in 4-bit.

External benchmark · DarijaCS · Prism & Probe

Character error on DarijaCS (lower is better)

Prism (a.k.a. ISMA)

  • fihrIA-ASR 1.7B
    6.05%
  • Voxtral Mini 4B, as published
    6.68%
  • fihrIA-ASR 0.6B
    7.90%
  • MoulSot
    10.27%

Probe (a.k.a. Darija in the Wild)

  • fihrIA-ASR 1.7B
    8.52%
  • Voxtral Mini 4B, as published
    9.69%
  • fihrIA-ASR 0.6B
    9.79%
  • MoulSot
    14.93%

Voxtral’s bars use the figures from its own chart. The other bars are our measurements at 16-bit precision, on the same clips and with the same protocol as DarijaCS.