Which open-weight text-inference LLMs were released in July–October 2026? For each: parameter count, active parameters, releasing lab, the lab's country, and whether it accepts image or audio inputs besides text — plus all available major benchmark scores, separating lab-run from external evaluations of the same benchmark target.
A live-web scan through 16 October 2026 surfaced 67 candidate open-weight text-generating model releases — 29 in July, 15 in August, 13 in September, and 10 in October (month-to-date through 15 October). This is a candidate census, not a guaranteed complete registry: 37 rows rest on a single source mention and are marked weakly corroborated; blank fields mean the sources did not state a value, never zero. Exact-model benchmark scores were recovered for 7 of the 67 models.
Each circle is one release, placed by date (x) and total parameter count (y, log scale); area scales with corroborating mentions. Colour is the lab's base country. 11 rows with no stated parameter count sit in the "params not stated" band at the bottom. Hover or tap a circle for lab, active parameters, modalities and source.
Country means the lab's headquarters, not training location. "Weak" marks single-mention candidates. "Multimodal" without enumerated inputs is kept as stated, unclear. Tencent Hy3 appears twice under near-identical rows (possible alias) and DeepSeek V4 Pro variants may overlap.
| Model | Date | Lab | Country | Params | Active | Image/audio input | Mentions | Source |
|---|
Blue badge = the releasing lab evaluated its own model; orange = external evaluator. Scores are comparable only when benchmark, target/variant, metric and conditions match; where both provenances tested the same model the rows sit together under the model. DeepSeek V4 Pro's 23 rows are all lab-run — breadth there does not imply independent validation. Benchmark coverage is incomplete: absence of a model here means no exact-model score was recovered, not that none exists.
Method: live-web scan completed 2026-10-16 covering open-weight text-generating model releases dated 2026-07-01 to 2026-10-15; 67 candidate release rows (October is month-to-date), 41 lab-to-country lookups, and 66 exact-model benchmark rows for 7 models. Rows are candidates, not a verified registry: 37 have a single corroborating mention, some names may be aliases (Tencent Hy3, DeepSeek V4 Pro variants), and several clusters come from release aggregators. Parameter counts are totals as reported; active parameters are MoE active counts and blank means not separately stated. Modality blanks mean unresolved, and open weights does not imply an open-source license. Benchmark scores are shown as published and compared only within identical benchmark targets and metrics. Evaluation notes were shortened for space.