Introducing bear-2-finance: A Query-Aware, Finance-Tuned Compression Model

bear-2-finance is a query-aware compression model for financial documents, now in private preview. Given a filing and a question, it returns only the relevant spans from the filing at a user-specified budget. On FinanceBench, bear-2-finance cuts filing context and input cost by 50% with no measurable accuracy loss.

Summary

SEC filings typically run 100,000–250,000 tokens, and the longest filing in our test set is 419,000 tokens long. If a query spans multiple filings, this can often exceed the maximum context length supported by large language models, requiring an overflow policy such as truncation. On FinanceBench, bear-2-finance cuts every filing to 50% of its original token count while matching full-document accuracy, and the end-to-end pipeline outperforms sending raw filings because compression lets oversized documents fit within the reader's context window. Readers hold their full-document accuracy at every budget from 90% kept down to 50%: Claude Haiku 4.5 scores 81.3 at 50% kept versus 81.1 on the full document; Gemini 3.5 Flash Lite scores 79.0 at 50% versus 79.3.³

Why filings need compression

SEC 10-K filings contain a great deal of irrelevant information and legal boilerplate, which cost money and prefill latency on every call. LLM readers have to attend past these tokens, which measurably degrades answers.¹ In reality, most single-hop and multi-hop queries can be found in a few segment footnotes, covenant terms, and line items, which typically add up to a few thousand tokens out of the full document.

In addition, uncompressed filings can often exceed the context window of production readers. Any pipeline that consumes larger or multiple filings must either truncate or separate inputs into multiple queries. Query-aware compression allows each sub-agent to use its own focused view of the filing rather than the full document.

How bear-2-finance works

bear-2-finance is a query-aware model, which means it can achieve stronger compression when provided with a focus statement alongside the primary document(s); the focus statement can be a user question, an instruction, or an agent provided sub-task. The user sets a retention budget between 10% to 90%, and bear-2-finance returns the relevant spans in plaintext format in under a second.² This is a straightforward API call with a response that can be fed directly to the reader model with no downstream changes required. Because the output is purely extractive, the compression stage ensures that all kept numbers, figures, dates, and defined terms survive verbatim, and every retained span exists at a definite location in the source document for auditing and citations. The model is optimized for financial text, but still functions with standard fidelity on out-of-domain and multilingual documents (similar to the base bear-2 compressor currently available on our API).

Evaluation setup

We measured on FinanceBench, which consists of 150 analyst questions over full SEC filings. The downstream reader models were Claude Haiku 4.5 and Gemini 3.5 Flash Lite. We trained five seeds of the bear-2-finance recipe and ran three independent reads per seed per condition.³ A GPT-5-mini judge model was used to score the outputs against the reference answers.

Cost and latency

These workloads cost 50% less and are answered with a 25 to 45% lower time to first token, compression time included.

Results

The production pipelines answer all 150 questions, including the ones about the 419,000-token filing. We scored three pipelines: query-aware compressing every filing to 50%; sending full documents and truncating oversized ones to fit the reader; and sending full documents with no overflow fallback. At 50% of the input tokens, bear-2-finance scores 81.3 end-to-end, versus 77.3 with truncation and 72.0 when oversized filings are allowed to fail (18 of the 150 FinanceBench questions use documents that exceed Claude Haiku 4.5's 200k token context window size).⁴ Both readers hold their full-document baseline at every budget from 90% kept down to 50%:

Comparison with truncation and retrieval

Extractive compression is not the only way to ensure filings fit in the context window: truncation, retrieval (BM25), and embeddings can also work. We compared bear-2-finance against them in two matched settings — first at a fixed context-window limit, then at a constant retention rate of 33% applied to every filing.

At a fixed context-window limit, where the only job is to make each filing fit, targeting 50% retention and falling back as needed to fit the context window stays ahead of every raw-document policy — including the truncation and embedding-retrieval fallbacks a production pipeline reaches for when a filing overflows. Holding the budget constant at 33% for every filing instead of the window tells the same story:

At an equal 33% budget, bear-2-finance outperforms both truncation and BM25 retrieval. And a standard retrieval stack also requires document chunking and embedding overhead, as well as persisting a vector store, compared to a simple one-off bear API call.

Notes

1. 67% retention rate compression performs better than uncompressed context, tested on GDPval.
2. Tested in production with 800k token inputs.
3. To measure performance, we trained five seeds of the bear-2-finance model and tested each seed with three independent reads. Tables and figures report the five-seed mean.
4. For the 419k token document, the 50% kept condition falls back to 33% kept.