Dimmed sections with a dashed outline are placeholders that show the design. Everything at full strength is real.
Tern-3: a 14B open-weights coding model that fits on one 24 GB card
Apache-2.0, training mix disclosed, eval harness in the repo. Real and useful; not, by most voters’ reckoning, new.
21 OCT 2026 · 06:10 ET
4 of 5 evidence links confirmed
Summary
Tern Labs released Tern-3, a 14-billion-parameter model tuned for code completion and repository-level edits. The weights are on a public hub under Apache-2.0. At 4-bit quantisation it runs on a single 24 GB consumer GPU with room for a 32k context.
The release includes the training mix down to the list of source repositories, and the evaluation harness used for the model card. Two of the three headline benchmark suites reproduce within a point; the third depends on a dataset that is still behind a request form.
Primary evidence
Checked against the claim, not the press releaseEditors’ note
An editor · 21 OctIt is what the card says. We pulled the weights, matched the checksums, and ran the published harness on a single 4090: two suites landed within a point of the claimed numbers.
Novel it is not. This is a careful refinement of a well-known recipe, and the voters have scored it accordingly. But careful refinements with honest licences are what people deploy, and that is the point of this list.
In the knowledge graph
Extracted from the evidence links and the item text · confirmed by an editor
Related
Comments
Commenting arrives with sign-in.
Pulled the weights on a 4090 and got the same throughput as the card. The 32k context fits, just.
Be careful with the third benchmark: the request form is real, and nobody I know has been approved yet.
We run it for code review bots. Good enough that the humans stopped skimming the diffs less carefully, which is the point.