Dragonscale's kniv v7: local fact extraction, 90x faster than GPT-6

Dragonscale, an AI company behind the Rustic AI initiative, published kniv v7 on 2 November 2026: a family of six open models that turn conversation text into entities and facts without calling a hosted LLM. The company says the largest are accurate enough to compete with a frontier model on extraction, at a fraction of the time and cost. This is the builder's own account, and the benchmark comparisons are its own.

What it does

Agent memory systems must pull facts out of what users say. Dragonscale says the usual choice is an LLM call per message, which is flexible but slow and billed, or a small tagger, which is cheap but finds only what it was trained on.

kniv reads for entity types and relations you describe in plain English, with no retraining. Each fact carries attributes: asserted, negated, hoped for or reported, plus holder and time. It reads up to a page at once, so a fact split across sentences by a pronoun can still be linked.

The models run from 55M parameters to 538M, are Apache-2.0, ship as ONNX files and need only onnxruntime, tokenizers and numpy. Dragonscale recommends the 221M "large" model as a start: within a point or two of the best on most tasks at about twice the speed.

The speed and cost claim

Dragonscale timed both systems on 60 user messages from the LongMemEval memory benchmark (median 141 words each), using a schema of 30 relations and 15 entity types. GPT-6 through its hosted API took a median 31.4 seconds per message, with 1,820 tokens in and 2,248 out. kniv's UltraBERT model took 0.35 seconds on an M2 Max laptop CPU; the DeBERTa variant took 0.53. The company puts the gap at about 90 times, with no per-message API fees and nothing leaving the machine.

The post gives no dollar figure for the GPT-6 calls and no server throughput, so cost per thousand messages is yours to work out.

The accuracy claim, and its limits

The strongest result is document-level relations. On the full Re-DocRED test set, with reference entities supplied to both systems, Dragonscale reports 0.733 F1 for kniv against 0.477 for GPT-6. Even when GPT-6 was handed every candidate entity pair, at 5.4 times the calls, it reached 0.545. The company says 75% of GPT-6's misses were pairs it never proposed.

On tagging tasks kniv's reported scores are higher by up to 4 points (0.870 against 0.849 on OntoNotes entities), but kniv was scored on full test sets and GPT-6 on 300-item samples, so Dragonscale says to read them as reported scores, not matched runs. Coreference was level. On unseen entity types kniv trails GLiNER-large (0.580 against 0.613).

A probe on 150 LongMemEval questions checked only whether the answer reached the extracted store: 0.69 of questions for entities and 0.48 for facts with UltraBERT, weakest on temporal and multi-session questions. No LLM answered any question, so this is not an end-to-end memory score.

Where Dragonscale says it breaks

• About 29% of gold facts are missed on Re-DocRED. Lowering thresholds finds more of them, with more false ones.
• Negation and reported speech are the weakest attributes: in the company's spot checks "I don't work there anymore" can come back as a positive fact.
• Facts are per window of up to a page; nothing links across windows or documents, which the post says is the memory layer's job.
• The smaller models return more low-margin facts and trail the teachers by 4 to 17 points on several tasks. English only.

What it means for a rollout

This is my reading, not Dragonscale's. If an agent writes memory from every message, extraction is a per-message cost and a latency step, and a local model removes both. The price is precision: filter on margin, treat negation as a signal to check, and expect a quarter or more of facts to be missed. Run your own transcripts and schema through both approaches and count the facts a person would keep.

Not in the post: independent replication, non-English results, or a cost for the hosted calls. Dragonscale is grading its own work.