Public story · 2026-08-04 · high

Intology's Locus agent beats Qwen3's human-tuned model

Intology's own benchmark variant, using thousands of H100 hours instead of the standard one, is what produced the win.

Why now: The result appears in the August 4 briefing, with Intology's own caveat about the benchmark attached.

Issue

2026-08-04

Confidence

high

Sources

Redaction

passed

Story

Intology's coding agent Locus post-trained a Qwen3 base model past Qwen3's own official Instruct checkpoint, per the company's blog.

On the actual PostTrainBench task, one H100 GPU and 10 hours across seven benchmarks, Locus scored 44.7%. The 51.6% figure that beats the official Qwen3-1.7B-Instruct model came from a different test. Intology built an expanded version of PostTrainBench itself, running thousands of H100 hours instead of one.

Intology calls this variant PostTrainBench+. It's the company's own extension of a third-party benchmark, not an entry on any external leaderboard. The blog post flags this directly: the headline comparison isn't apples to apples with how PostTrainBench was designed to be run.

The real number

Watch the 44.7%, not the 51.6%. That's Locus's score under the constraint the benchmark actually enforces: one H100, ten hours. It's the number that means something if another agent tries the same task on the same budget. The 51.6% only proves that more compute helps, which nobody needed an AI agent to demonstrate.

Source trail

Intology

Entities

H100 Instruct Intology Locus Note PostTrainBench Qwen3

Claim evidence

Intology's Locus agent beats Qwen3's human-tuned model
Source

Provenance

Canonical issue: 2026-08-04
AI generated: yes
Story unit: 2026-08-04-intology-s-locus-post-trains-qwen3-base-models-past-the-official-human-tuned-instruct-checkpoint
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Intology's Locus agent beats Qwen3's human-tuned model

Story

The real number

Related stories

PostTrainBench: Agents Autonomously Post-Train LLMs — But Reward Hack When Unsupervised

PostTrainBench: Claude Opus 4.6 Leads at Agent-Trained-Agent Tasks — and Discovers Reward Hacking.

PostTrainBench: Can Agents Automate LLM Post-Training?

Codex on GPT-5.5 tops Terminal-Bench 2.1 at 83.4%, but the top Claude config is export-banned.

A 27B Model Just Matched the 400B One. It Fits on Your Laptop.

impeccable hits 50,392 stars shipping design taste as an installable skill.

First Benchmark Catches Coding Agents Acting Outside Their Authorization. The Numbers Are Uncomfortable.

Garry Tan's GStack Hits 10K Stars in 48 Hours: The Thesis That Structured Prompts Are All You Need

Source trail

Entities

Claim evidence

Provenance

Intology's Locus agent beats Qwen3's human-tuned model

Story

The real number

Related stories

PostTrainBench: Agents Autonomously Post-Train LLMs — But Reward Hack When Unsupervised

PostTrainBench: Claude Opus 4.6 Leads at Agent-Trained-Agent Tasks — and Discovers Reward Hacking.

PostTrainBench: Can Agents Automate LLM Post-Training?

Codex on GPT-5.5 tops Terminal-Bench 2.1 at 83.4%, but the top Claude config is export-banned.

A 27B Model Just Matched the 400B One. It Fits on Your Laptop.

impeccable hits 50,392 stars shipping design taste as an installable skill.

First Benchmark Catches Coding Agents Acting Outside Their Authorization. The Numbers Are Uncomfortable.

Garry Tan's GStack Hits 10K Stars in 48 Hours: The Thesis That Structured Prompts Are All You Need

Source trail

Entities

Claim evidence

Provenance