Thomson Reuters’ Thomson model cost $40M. The outside test is still missing
Thomson Reuters spent $40 million building a legal-focused model. Its first product use is clear, but the public benchmark evidence remains internal.
Frontier models, comparisons, and how to choose what fits your work.
Thomson Reuters spent $40 million building a legal-focused model. Its first product use is clear, but the public benchmark evidence remains internal.
NVIDIA says AVO completed all 183 public ARC-AGI-3 levels with 100 RHAE. The architecture is useful evidence; private-set generalization is still untested.
LLMRouter compares more than 16 routing methods across quality and cost. Here is how to test the route, overhead, failure rate, and final bill.
NVIDIA's Nemotron 3.5 Lightning activates 3B of 30B parameters, but the BF16 weights are about 66 GB and one-million-token context needs serious hardware.
Meta Muse Glimmer offers a 17 GB package for 24 GB devices. Builders still need to test runtime support, context memory, and tool-loop reliability.
SHE turns unsafe agent trajectories into bounded safety-harness edits. Its held-out test matters more than the self-editing demo.
A new Text-to-SQL study separates exact replay from held-out transfer and reports that verified repairs improved later first-attempt accuracy.
Mistral Shieldstral accepts a natural-language safety policy at inference time. Builders still have to calibrate thresholds, cost, and human review.
An AutoML benchmark looked decisive until its authors separated model selection from the test set and enforced equal wall-clock limits.
P-Bench tests whether AI agents choose an appropriate statistical method before running R code. Fisher-R1 improves reported results, but its public artifacts are incomplete.
FinRank tests whether financial QA systems retrieve evidence from the correct company, filing period, and disclosure context, not merely a plausible passage.
PAST-Bench tests agent memory across clean sessions. A new TEPA update shows why useful memory also needs explicit validity, revocation, and an audit trail.