Is it good enough?
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.
- AssessUse & modify + Transparency
- ImplementData control + Reliability
- UseReliability
- SupportTransparency
Capabilities & modalities
Text only. The post-training data covers maths, science, coding, general chat, RAG and translation (ev-card-instruct, ev-posttrain-data). The publisher's figure shows strong instruction following, with the best ifeval and ifeval-fr scores of the five models compared (ev-bench).
Maths and general chat trail Mistral-Small-3.2 in the same figure. The card says code and maths need additional training. All numbers are the publisher's.
Context window & long-context behaviour
The configuration sets max_position_embeddings to 131,072. The base model reached that window through two 50-billion-token extension phases, to 32,768 and then 131,072 tokens (ev-card-base). Post-training used 16,384-token sequences, and the card says the Instruct model was trained on shorter sequences than its base (ev-card-instruct).
The base card reports RULER results. No long-context result is published for the Instruct models. Test your own long prompts before relying on the window.
Prompt format & chat template
The template follows the ChatML layout, which the card says is inspired by Qwen3-1.7B. A default system prompt reads "You are a helpful AI assistant named Luciole, trained by LINAGORA and OpenLLM France." Turns look like this (ev-card-instruct):
<|im_start|>system
You are a helpful AI assistant named Luciole, trained by LINAGORA and OpenLLM France.<|im_end|>
<|im_start|>user
Who was Molière?<|im_end|>
<|im_start|>assistant
Use tokenizer.apply_chat_template(..., add_generation_prompt=True), which reads chat_template.jinja. The end token is <|im_end|> (id 261 in generation_config.json).
Language coverage
The Instruct cards list French and English. Post-training data is almost entirely English, with French from Nemotron multilingual maths (80K in the thinking mix), Paradocs and Croissant translation sets, and about 2,000 hard-coded identity prompts. The safety pairs cover English, French, German, Spanish and Italian (ev-posttrain-data).
The base card lists the pretraining shares: English 41.9%, French 30.4%, German 3.8%, Spanish 3.5%, Italian 1.9%, Portuguese 1.3%, Dutch 1.0%, Arabic 0.5%, and 0.4% for regional languages and French creoles (ev-card-base). The Base model was evaluated on German, Spanish, Italian, Portuguese, Dutch and Arabic. The Instruct card makes no claim for those languages.
Function / tool calling
vLLM's hermes parser turns tool calls into message.tool_calls. The training mix includes xlam, hermes, When2Call and smolagents data (ev-posttrain-data). The card warns that tool use is work in progress and links BFCL v4 charts on the 1B repo (ev-card-instruct). Treat tool calling as unproven on your tools until you test it.
How this scores
The ownership factor this domain covers, drawn from the one entry record.
ReliabilityIs it reliable and good enough for the job?
WeakOperational readiness is 3, and safety is 1 because no safety evaluation, guard model or acceptable-use policy is published. The Thinking card states that safety alignment can still leave unsafe output (ev-card-instruct, ev-card-thinking).
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.