Check an agreement. Explain a charge. Find what is still missing. AMBER connects your question with the relevant document passages; the model helps turn them into an answer you can review.
Start with the way your team works, the information it handles and the equipment it has.
Gemini · cloud
For a responsive working conversation.
The answer model runs with the provider. Your team uses AMBER in a browser, without a local machine sized to run the model.
Selected document passages are sent to the provider to prepare the answer.
Observed response times
2–4 seconds for simple requests; 17–36 seconds for more involved questions.
What to budget for
Model usage charges, application hosting and maintenance.
Qwen · local
For analysis on your own equipment.
Our trials show that a local model can work with agreements, AML materials and legal documents. This route suits teams that want the answer model on their own infrastructure.
The local route was tested in research. We agree the complete data flow and prepare the deployment for each environment.
Observed response times
Around 70 seconds typically in our public-document trial.
What to budget for
Suitable hardware, electricity and maintenance, alongside the application.
Times come from different studies and tasks. They illustrate the experience, not a direct speed comparison or a delivery promise.
What we observed.
Six document tasks. The same supplied evidence. Five local models, with Gemini as a cloud reference.
Requirements met in the six-task comparison
Gemini 3.8 Flash high34 / 34
Qwen 3.6-35B-A3B Q430 / 34
Ornith 1.5-35B-A3B Q429 / 34
Gemma 4-26B-A4B Q426 / 34
Nemotron 3.5 Lightning Q419 / 34
GLM 4.7 Flash Q412 / 34
34 individual checks across six prepared tasks. These scores are not the percentage of fully correct answers.
Gemini met all 34 checks in this comparison.
Qwen met 30. It handled useful calculations and interpretation, but needed closer review of sources and answer completeness.
Check important conclusions and their sources with every model.
Which agreement applies, and what changed in the amendment?
02
Explain a charge
What remains payable after the credit already applied?
03
Review the evidence
What do these ownership records establish, and what is still missing?
About these observations
Research observations · September 2026
The six-task comparison used prepared documents and the same drafting instructions. Gemini 3.8 Flash high met 34/34 checks; Qwen3.6-35B-A3B, with thinking off, met 30/34. Model settings differed. The documents were supplied directly: this comparison measures the answers, not a gain attributable to AMBER search.
Task-by-task results
Model
Requirements met
Complete cases
Median time
What to consider
Gemini 3.8 Flash highCloud reference
34 / 34
6 / 6
Not comparable
Met all measured requirements in this sample; important conclusions still need review.
Qwen 3.6-35B-A3B Q4
30 / 34
3 / 6
56.3 s
Highest local score in this sample; omitted required details and the historical comparison.
Ornith 1.5-35B-A3B Q4
29 / 34
2 / 6
35.9 s
Close to Qwen, with shorter responses; double-counted a credit and overstated one ownership share.
Gemma 4-26B-A4B Q4
26 / 34
1 / 6
22.5 s
Shortest local median here; double-counted a credit and omitted required distinctions.
Nemotron 3.5 Lightning Q4
19 / 34
0 / 6
29.5 s
Errors in amounts, party identification and source references.
GLM 4.7 Flash Q4
12 / 34
0 / 6
23.0 s · attempts
Five final answers delivered; numerical errors and one answer cut off at the output limit.
A complete case meets every task requirement and the application’s source-reference and comparison rules. Zero complete cases does not mean every statement was wrong. These six prepared tasks are observations, not a general model ranking.
Local medians cover drafting plus source checking on the same six supplied contexts, excluding retrieval, import and scan processing. GLM delivered five of six final answers; its 23.0-second median includes the failed 194.9-second attempt. The separate Vertex times above are not timings for this six-task Gemini result.
Primary, strict scores are shown. Alternative interpretations of individual requirements give Ornith 30/34 and 3/6 complete cases, Nemotron 20/34, or GLM 13/34. These alternatives do not replace the primary assessment.
Where each model succeeded or stumbled
Task
Gemini 3.8 Flash high
Qwen 3.6-35B-A3B Q4
Ornith 1.5-35B-A3B Q4
Gemma 4-26B-A4B Q4
Nemotron 3.5 Lightning Q4
GLM 4.7 Flash Q4
Freight charge correction
6/6All six checks met
5/6Correct balance; unsupported claim about the rate’s origin
4/6596 instead of 620: credit deducted twice
4/6596 instead of 620: credit deducted twice
1/6Free waiting time omitted
0/6266 instead of 140 for waiting
Earlier contract terms
6/6All six checks met
5/6Correct terms; required comparison not accepted
5/6Correct terms; required comparison missing
6/6All six checks met
4/6Retroactive effect denied despite the source
5/6Comparison order reversed
Laboratory release
5/5All five checks met
5/5All five checks met
5/5All five checks met
3/5Other instrument and lot not explicitly excluded
3/5Same omissions; English instead of German
3/5Other instrument and lot not explicitly excluded
Ownership and control
6/6All six checks met
6/6All six checks met
5/672% instead of 36% for one owner
5/6No explicit request for full identifiers
6/6Six criteria met; source declaration missing
2/625.2% instead of 36%; other shares omitted
Pleading review
6/6All six checks met
4/6A2 section 2 omitted; supplied court rejected
5/6Court identified, but source and missing-field distinction confused
4/6A2 section 2 and missing attachment link omitted
3/6Wrong company and incomplete clause reference
0/6No final answer: output limit reached
Spend and coverage
5/5All five checks met
5/5All five checks met
5/5All five checks met
4/5Earlier-record completeness caveat omitted
2/531,600 instead of 29,600: credit not subtracted
2/522,000 instead of 29,600
Scores count individual requirements. Notes give selected examples; a full score does not guarantee flawless wording or acceptance by every application check.
Cloud times cover eight successful simple or conversational requests (2.12–3.53 seconds) and four complex requests (16.84–36.05 seconds) through Google Vertex on small prepared document sets. Local times come from 18 Polish questions over eight public AML, customs and legal documents: median 69.77 seconds, range 48.28–96.73 seconds. These are separate studies and application versions; import and scan processing are excluded. Concurrent use was not measured.
In the public-document study, Qwen met 50/54 content checks; 10/18 answers met all content, language and source-support requirements (11/18 under an independent review). The trial did not meet its pre-set quality and response-time targets. Local use therefore remains a route to assess in a pilot, with human review.
Try the question that matters to you.
Bring a small, agreed sample of documents. Together we assess the answer, its sources and the time it takes before choosing the deployment.