Needs rechecking. These attributed findings were checked against search excerpts, not hands-on tests. Identity is linked to a primary reference; aliases and renames may still need resolution. Availability and usefulness are not independently assessed. The observed date is not a launch date.
What the sources describe
What it is
Sevren describes Heimr 570M as experimental, built from Mamba-3 layers plus its lab's Heimr attention layers, with 2.7B total and roughly 570M active parameters per token. Source ↗
The company says Heimr 570M was trained to a 64K context and decodes 2 to 8× faster than comparable open models on long contexts. Source ↗
This is a short discovery note. The available evidence does not yet support a complete practical guide.
Research history
Dated observations, not a claim that the product changed each time. Identical evidence does not create a new entry. Later checks can update the last-checked date.
View this observation
Sevren describes Heimr 570M as experimental, built from Mamba-3 layers plus its lab's Heimr attention layers, with 2.7B total and roughly 570M active parameters per token. Source ↗
The company says Heimr 570M was trained to a 64K context and decodes 2 to 8× faster than comparable open models on long contexts. Source ↗
How this was found
Initial broad query: new open weight local AI models inference datasets releases independent hands-on experiences since 2026-09-28
Model: deepseek/deepseek-v4.1-flash · Policy: web-exploration-v2. Search excerpts are held privately for seven days; source links, evidence hashes and our reviewed paraphrases remain. Linked pages may change.