Hugo Vergnes reports training a 3.8B language model to a 0.384 CORE score for $998, potentially making small-model training at that measured quality accessible on a roughly $1,000 compute budget.

state: seedheat: lowuncertainty: highnovelscott: lowopen-models training-costHugo Vergnes

What is this?

A write-up attributed to Hugo Vergnes is titled “Training a 3.8B LLM to 0.384 CORE for $998,” reporting a low-cost language-model training result. None of the supplied search-result snippets covers Vergnes or this experiment; they discuss other small models, so the search summary’s repetition does not independently verify the claim. The supplied material does not establish what CORE measures, whether this was training from scratch or further training, what the $998 includes, or whether weights and a reproducible recipe are available.

Why it matters to Scott

Scott’s Salesforce fine-tuning data factory is adjacent, but the supplied claim establishes neither a usable training recipe for that work nor task-level quality or inference economics that would alter his Model Barbell choices. CORE’s meaning, the training regime and the $998 cost scope remain unspecified; the radar tracks related low-cost training experiments, but no supplied page tracks this Vergnes result.

radar:concept.small-model-trainingradar:concept.training-efficiencyradar:prime-intellect-nanogpt-speedrunradar:bananamind-2-pro-consumer-gpu-training

queries asked of Scott's wikis

small-model training economics compute budgets
benchmark validity capability claims cost accounting
open weights reproducible training model ownership
domain-specific models fine-tuning versus RAG
self-hosted models agent workloads quality thresholds

Measured heat

now 0 pts/hpeak 0 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 758h

points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion

How the heat travelled

09-10 02:22 (minted)	⭐ origin echo-reconstructed	The linked write-up is titled “Training a 3.8B LLM to 0.384 CORE for $998.” Hugo Vergnes on blog (echo) · attributed from hn.story.49637435 · published time unknown	—
09-10 02:04	first on hacker news · published · lag ?	Training a 3.8B LLM to 0.384 CORE for $998 – Hugo Vergnes Anon84	—
09-10 02:04	amplified on hacker news 👑	hn.story.49637435 Anon84	peak 121 · 21 comments · 100% of case engagement
09-10 02:21	our radar first saw it · lag ?	discovery anchor: hn.story.49637435	—

pace: p71 vs 519 stories at the 720h mark (now 758h old) — ahead of antigravity-boost-reasoning-control (1.0x), behind engrim-local-cli-memory (1.0x)

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟧 hn	Training a 3.8B LLM to 0.384 CORE for $998 – Hugo Vergnes	Anon84	121	20
🟧 echo.blog ⭐	The linked write-up is titled “Training a 3.8B LLM to 0.384 CORE for $998.”	Hugo Vergnes	—	—

Interpretation history

2026-09-10T08:27:40Z
The refreshed discussion adds enthusiasm and hypothetical follow-up experiments, not a replication, released implementation or clarification of training economics. The $998 result remains a title-level report from one reporting chain, with no new evidence that it offers Scott a usable small-model training option.
2026-09-10T02:32:41Z
No substantive evidence has arrived: the HN link and reconstructed write-up title remain one unverified reporting chain, not corroboration of usable training economics. The supported claim is $998, not the case title’s truncated $9; training regime, cost scope, benchmark meaning and reproducibility remain unspecified.
2026-09-10T02:25:41Z
grounded: novel/low — Scott’s Salesforce fine-tuning data factory is adjacent, but the supplied claim establishes neither a usable training recipe for that work nor task-level qualit
2026-09-10T02:22:58Z
case created — A specific cost-and-quality result with an author-owned write-up warrants tracking, while cost accounting and reproducibility remain unestablished.

Decision trace

10-04 19:26review_dormantscheduled targets exhausted or 28 quiet days
10-04 19:26drop_targetsquiet through full ladder or over cap 8
09-12 05:36review_screenThe added comment expresses interest in a future open-source training project but provides no release, implementation, replication, or new training-cost evidence; the other discussion remains unchange
09-11 02:21sensor_dirtyengagement_update
09-10 23:21sensor_dirtyengagement_update
09-10 22:21sensor_dirtyengagement_update
09-10 21:21sensor_dirtyengagement_update
09-10 20:21sensor_dirtyengagement_update
09-10 19:21sensor_dirtyengagement_update
09-10 18:27repriceThe refreshed discussion adds enthusiasm and hypothetical follow-up experiments, not a replication, released implementation or clarification of training economics. The $998 result remains a title-leve
09-10 18:27alert_silentThe new comments establish no consequential release or technical result; proposed architecture changes are suggestions, not implementations. There is no new decision for Scott that would make the next
09-10 18:27alert_routeThe new comments establish no consequential release or technical result; proposed architecture changes are suggestions, not implementations. There is no new decision for Scott that would make the next
09-10 18:21sensor_dirtycomment_update
09-10 17:21sensor_dirtyengagement_update
09-10 16:21sensor_dirtyengagement_update
09-10 15:21sensor_dirtycomment_update
09-10 14:21sensor_dirtyengagement_update
09-10 12:32repriceNo substantive evidence has arrived: the HN link and reconstructed write-up title remain one unverified reporting chain, not corroboration of usable training economics. The supported claim is $998, no
09-10 12:32alert_silentThere is no new consequential delta or established recipe, weights release or cost detail that changes Scott’s decisions. The existing title-level claim can wait for the briefing; the engagement chang
09-10 12:32alert_routeThere is no new consequential delta or established recipe, weights release or cost detail that changes Scott’s decisions. The existing title-level claim can wait for the briefing; the engagement chang
09-10 12:31alert_silentA specific low-budget training claim is worth retaining, but the supplied evidence contains only the write-up title. It establishes neither the training regime and cost scope nor what the CORE score m
09-10 12:31surface_candidateA specific low-budget training claim is worth retaining, but the supplied evidence contains only the write-up title. It establishes neither the training regime and cost scope nor what the CORE score m
09-10 12:31alert_routeA specific low-budget training claim is worth retaining, but the supplied evidence contains only the write-up title. It establishes neither the training regime and cost scope nor what the CORE score m
09-10 12:25groundScott’s Salesforce fine-tuning data factory is adjacent, but the supplied claim establishes neither a usable training recipe for that work nor task-level quality or inference economics that would alte
09-10 12:22createA specific cost-and-quality result with an author-owned write-up warrants tracking, while cost accounting and reproducibility remain unestablished.