Blue

Every model knows who Napoleon was. We care about the other seven million people.

Blue is Red Kit's open line: a model you run on your own hardware that can recall genuinely obscure facts, backed by an encyclopedia index that lives on your disk instead of someone's API. Weights, training data and method all published.

The problem worth solving

Language models are excellent on famous subjects and quietly terrible on rare ones — and they are most confident exactly where they are least reliable. That gap is invisible on the benchmarks the industry quotes, because those benchmarks are dominated by well-covered material.

It matters most for the people who cannot use a frontier API in the first place: anyone working offline, on private data, under a records policy, or on the wrong side of a network. They get handed the smallest models, which are the ones that hallucinate hardest about the long tail.

Blue's question: how much rare, checkable knowledge can you actually get out of a model that runs on hardware you own — and can you prove the number honestly?

The library is built

Blue is grounded on a full encyclopedia snapshot, processed and indexed so a model can search it locally at answer time. That part is done.

7.6Marticles, extracted and cleaned

53Msearchable passages, 7.6 billion words

$65total rented-GPU cost to build the whole index

Built on a rented fleet across four data centres, with zero data errors across 86 shards. How well it retrieves is the next thing we measure.

We picked the base model by testing nine of them

Rather than take a leaderboard's word for it, we ran nine local models plus a frontier reference through a thousand deliberately hard questions each, with no library to consult — pure recall, machine-graded.

Model	Rare facts	Popular facts
Qwen3.6-35B-A3B mixture-of-experts	16.5%	28.3%
Qwen3.5-27B dense	10.8%	26.5%
Qwen3.6-27B dense	9.0%	25.8%
Gemma 4 31B	7.4%	18.1%
Qwen3.5-9B	5.3%	17.7%
Bonsai 27B 1.58-bit	4.3%	17.2%

Six of the nine shown. Percentages are answers judged correct on a thousand questions per set.

The winner holds 35 billion parameters but only activates about 3 billion per answer — so it carries the most retrievable rare knowledge in the field while running several times faster than the dense models it beat. On the behavioural tie-breaker it finished close to the runner-up, and a rule we had written down in advance settled it on speed.

Two findings worth publishing on their own

Extreme compression destroys rare facts specifically. A 1.58-bit model keeps its general ability but drops to 4.3% on rare recall against 9–11% for its less-compressed relatives. The industry quotes a single "retention" figure for compression; the damage is concentrated in exactly the knowledge people compress models to keep.
Benchmark generations can move backwards on knowledge. One newer model posts a higher score on the standard reasoning benchmark than its predecessor and does worse on rare facts. If you are choosing a base model for a knowledge product, the popular proxies will actively mislead you.

Where Blue is

✓ Encyclopedia corpus extracted and chunked 7.6M articles → 53M passagesdone

✓ Search index built and hosted locally 86 shards, four data centres, zero data errorsdone

✓ Base model selected on measured evidence nine local models plus a frontier reference, thousand-question setsdone

▶ Train Blue v1 and measure what retrieval adds the closed-book control is already measured; the comparison is the experimentnext

· Publish weights, training data and method all three, not just the weightsplanned

What Blue inherits from Red

Blue has not trained yet, which means it starts with something better than a clean slate: a full campaign's worth of expensive lessons from the commercial line, written down and transferable.

Measure the grader before believing a null result. Red reported a run of "no significant change" findings before anyone asked whether the measurement could resolve a change at all. Blue's binary gates get a noise floor from day one.
A single average hides opposite-signed effects. Red found training made one subgroup better and another worse while the headline number barely moved. Blue's rare/popular split is declared up front, and reported separately, always.
Screen the training data with a model, not a keyword search. On Red, hand-searching found 33 problems and reading everything found 533.
A permissive licence says nothing about whether the content is right. Date the snapshot; check the claims.