Olala-v1.0-summarise

Olala-v1.0-summarise is a document summarization model, fine-tuned from Olala-v1.0-mini-7A1B. It reasons before it writes the summary.

  • Many document types: news, dialogues, forum posts, earnings calls, bills, government reports and scientific papers.
  • Structured and faithful: trained to write structured summaries (sections or labelled bullets) that stay faithful to the source, and to follow the requested form ("in one sentence", "as bullet points").
  • Concise: its summaries are shorter than those of 16 of the 17 compared models.

Model details


Model Description
Olala-v1.0-mini-7A1B-base Pre-trained base model (12.8T tokens)
Olala-v1.0-mini-7A1B-SFT After supervised fine-tuning, before RL
Olala-v1.0-mini-7A1B Final post-trained model: reasoning, chat and tool use
Olala-v1.0-summarise Specialized for document summarization
Olala-v1.0-translate Specialized for machine translation, including financial documents
Property Specification
Checkpoint Fine-tuned from Olala-v1.0-mini-7A1B for summarization
Parameters 6.8B total, about 1B active per token
Architecture 36 layers: 29 Mamba-3 and 7 attention layers, mixture-of-experts with 256 routed experts (6 active)
Context length 32K tokens (input + output)
Languages English (training documents are English except EUR-Lex)
License Apache 2.0

Inference


Sampling parameters: temperature=0.8, top_p=0.9, top_k=40, reasoning_effort="medium". Cap max_tokens: 99% of generations finish under 2K tokens, and 4K to 8K is a safe cap.

vLLM: always serve this model with --mamba-ssm-cache-dtype float32. vLLM's default keeps the Mamba recurrent state in bf16, and long generations then degrade into repetition loops.

vLLM quick start (vLLM fork, installed as for Olala-v1.0-mini-7A1B):

vllm serve OVHaiLLM/Olala-v1.0-summarise \
    --trust-remote-code \
    --mamba-ssm-cache-dtype float32 \
    --reasoning-parser olala \
    --max-model-len 32688 --port 8000
curl -s http://localhost:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "OVHaiLLM/Olala-v1.0-summarise",
  "messages": [{"role": "user", "content": "Write an executive summary of the following government report: why the work was done, what was found, and what is recommended.\n\n<report text>"}],
  "temperature": 0.8, "top_p": 0.9, "top_k": 40, "max_tokens": 8192,
  "chat_template_kwargs": {"reasoning_effort": "medium"}
}'

The summary is in choices[0].message.content and the reasoning in reasoning_content.

Transformers quick start:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "OVHaiLLM/Olala-v1.0-summarise"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True, dtype="bfloat16",
                                             device_map="cuda")  # or "cpu" without a GPU (slow)

document = open("article.txt").read()
messages = [{"role": "user", "content": "Summarize the following news article in one sentence.\n\n" + document}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True, reasoning_effort="medium",
                                 return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=4096, use_cache=True,
                     do_sample=True, temperature=0.8, top_p=0.9, top_k=40)
print(tok.decode(out[0, inputs["input_ids"].shape[1]:], skip_special_tokens=False))

Pass use_cache=True (config.json sets use_cache: false) and generate one prompt at a time.

Prompting

Put the instruction first, then a blank line, then the document. State the form you want ("in one sentence", "as short bullet points", "in one or two paragraphs"). With an open instruction ("Summarize the following paper."), the model tends to write a structured summary with sections.

reasoning_effort can be disabled, low, medium (default) or high. It barely changes ROUGE; medium is the setting all the numbers on this card use.

Chat template

The model uses the general Olala 1.0 mini chat template (chat_template.jinja, the same as Olala-v1.0-mini-7A1B). It writes its reasoning in an analysis channel, then the summary in a final channel:

<|im_start|>assistant<|channel_start|>analysis<|content|>...reasoning...<|channel_end|><|channel_start|>final<|content|>...summary...<|channel_end|><|im_end|>

Use tokenizer.apply_chat_template() with add_generation_prompt=True to format messages automatically, and set the reasoning effort with reasoning_effort=... (in vLLM, only inside chat_template_kwargs: "chat_template_kwargs": {"reasoning_effort": "medium"}, not the top-level reasoning_effort parameter). With the vLLM reasoning parser, the summary is in content and the reasoning in reasoning_content.

Performance


Benchmarks

Local summarization suite over 11 public test sets: up to 1,000 documents each (seed 42), documents cut at 12K words, one-turn prompt <instruction>\n\n<document>. ROUGE-Lsum F1 times 100 (stemmed), plain average over tasks. Length ratio = summary words / reference words.

Test set (ROUGE-Lsum) Olala
v1.0-summarise

7B-A1B
Falcon
H1-7B-Instruct
7B
Falcon
H1-3B-Instruct
3B
Mistral
7B-Instruct-v0.3
7B
NVIDIA
Nemotron-3-Nano-4B
4B
Granite
4.1-8B
8B
Qwen3.5-4B
4B
Gemma
4-E2B-it
2B effective
LFM2-2.6B
2.6B
Qwen3
4B-Instruct-2507
4B
LFM2.5
1.2B-Instruct
1.2B
Qwen3.5-2B
2B
SmolLM3-3B
3B
OLMoE
1B-7B-0924-Instruct
7B-A1B
LFM2-8B-A1B
8B-A1B
Gemma
3-4B-it
4B
LFM2.5
8B-A1B
8B-A1B
LFM3-1.7B
1.7B
News
CNN/DailyMail33.033.534.131.831.729.929.230.629.029.532.929.832.529.128.927.128.228.0
XSum16.521.822.418.316.718.418.117.516.918.016.916.316.716.916.418.615.714.8
Multi-News36.337.237.638.636.637.637.034.137.036.933.937.039.136.433.736.132.534.6
Dialogue
SAMSum33.432.831.532.533.133.128.832.230.132.929.429.633.130.628.928.929.630.1
DialogSum28.126.524.424.727.127.622.723.324.326.423.224.725.021.324.221.624.321.5
Forum posts
TL;DR21.225.218.518.718.117.916.619.016.218.218.215.617.921.716.317.515.715.7
Earnings calls
ECTSum15.915.415.014.513.712.415.813.213.111.513.413.613.29.412.113.112.511.3
Government and legal
BillSum38.840.340.044.537.337.036.037.241.031.743.635.038.838.641.035.235.732.9
GovReport43.049.448.049.647.451.150.949.646.549.144.249.951.445.842.745.741.444.3
Science
arXiv37.437.838.937.537.032.838.935.636.235.135.434.225.933.636.035.534.733.1
PubMed40.543.543.642.640.040.441.441.139.740.537.640.531.237.438.038.436.136.1
Average31.2733.0432.1732.1430.7830.7630.4830.3129.9929.9829.8729.6629.5229.1728.9428.8827.8627.47
Length ratio
summary / reference words
2.322.692.693.122.633.783.193.312.903.701.893.434.414.352.743.232.673.67
Evaluation details
  • This model: reasoning_effort=medium, T 0.8 / top_p 0.9 / top_k 40, 32K-token output cap, vLLM 0.26 (fork) with --mamba-ssm-cache-dtype float32. Other models: greedy, no thinking, 4K-token cap.
  • Summaries are shorter than most other models' (length ratio 2.3; the others range from 1.9 to 4.4). ROUGE rewards length up to a point, so part of the gap to the larger baselines is length, not content.
  • ROUGE does not reward the DPO goal (structured, faithful summaries); its effect shows in the judge scores more than here.
  • 6 of 10,287 generations (0.06%, earnings calls and bills) fell into a repetition loop in the reasoning and hit the 32K cap without a summary; they count as empty answers (ROUGE 0).
Effect of the temperature

Lowering the temperature does not raise ROUGE but makes reasoning loops more frequent. Use 0.8 and cap max_tokens, so that a rare loop costs little.

Temperature 0.8 0.6 0.4 0.2
ROUGE-Lsum average 31.07 31.19 31.22 31.13
Generations stuck in a loop, all test sets 0.04% 0.15% 0.30% 0.48%
Generations stuck in a loop, earnings calls only (ECTSum) 0.8% 1.8% 3.6% 4.4%

Training


Starting from Olala-v1.0-mini-7A1B, three stages:

  1. Warm-start SFT: one epoch on about 27,000 of the model's own best-of-4 summaries, with reasoning (rejection sampling).
  2. Reinforcement learning (GRPO): rewards from an LLM judge (faithfulness, coverage, conciseness, coherence, instruction following) mixed with a deterministic rubric (ROUGE against the reference, length, format, language). Training documents: news, government reports, EU legislation and financial filings.
  3. Preference optimization (DPO): about 1,900 preference pairs on scientific, medical, legislative and government documents. Chosen answers are structured and rated well on every judge criterion; rejected answers are either prose or structured but unfaithful.

Uses and limitations


Olala-v1.0-summarise is built for summarizing documents. The limitations of Olala-v1.0-mini-7A1B apply. In addition:

  • Summaries can omit or distort facts. Check figures, names and conclusions against the source before you rely on them.
  • Summarization only. For chat and tool calling, use Olala-v1.0-mini-7A1B.

Before you deploy: read the public summary of training content and our copyright policy.

License


Licence Apache License, Version 2.0 (SPDX: Apache-2.0)
Licensor OVH LLM SAS, trading as OVHai LLM, 2 rue Kellermann, 59100 Roubaix, France
Copyright Copyright 2026 OVH LLM SAS
Release date 19 November 2026
Distribution Open weights
Notices NOTICE: keep it when you redistribute the model or a derivative. It includes the attribution of the Qwen tokenizer (Copyright 2024 Alibaba Cloud, Apache 2.0)
Training content (EU AI Act) Public summary of training content, on the European Commission template (Art. 53(1)(d))

Citation


@techreport{ovhaillm2026olala,
  title       = {Olala v1.0 mini Technical Report},
  author      = {{OVHai LLM Research Team}},
  institution = {OVHcloud},
  year        = {2026},
  url         = {https://huggingface.co/OVHaiLLM/Olala-v1.0-mini-7A1B/blob/main/Olala-v1.0-mini.pdf},
}
Downloads last month
68
Safetensors
Model size
7B params
Tensor type
F32
ยท
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for OVHaiLLM/Olala-v1.0-summarise

Finetuned
(2)
this model
Quantizations
1 model

Collection including OVHaiLLM/Olala-v1.0-summarise