Data
Open-Model Tracker
One row per major open-weight model: publisher, license, parameter counts, and where it can be run — self-hosted, via the official API, or across hosted providers. Also available as JSON.
The Aug 2026 open-weight frontier
Seed data for the Aug 2026 open-weight frontier, compiled from DeAI's launch-sprint sourcing. Fields we could not confirm are null and render as "—", not estimated. Per-provider pricing columns populate as the weekly Price Index sweep runs (see /prices); nothing here should be read as a live price.Numbers we could not confirm are shown as “—”, never estimated. See the full sampling approach on methodology.
| Model | Publisher | License | Params (total / active) | Context | Where to run |
|---|---|---|---|---|---|
| Kimi K3 Parameter counts and license terms per Moonshot's own reporting; commercial-use caveats apply — read the license before deploying. | Moonshot AI | Custom (Moonshot) | 2.8T / 104B | — | Self-host (large-scale infra required) · Official API · Hosted providers — pricing pending weekly sweep |
| Kimi K2.5 Recorded on the Price Index as a list-price anchor (~$0.60 / $3.00 per M tokens, as of 2026-08-20); license and parameter counts unconfirmed as of seed date. | Moonshot AI | — | — | — | Official API · Hosted providers — pricing pending weekly sweep |
| DeepSeek V4 Pro License and parameter counts unconfirmed as of seed date — verify at draft time. | DeepSeek | — | — | — | Official API · Hosted providers — pricing pending weekly sweep |
| DeepSeek V4 Flash Recorded on the Price Index as a list-price anchor ($0.14 / $0.28 per M tokens, official API, as of 2026-08-20); license and parameter counts unconfirmed as of seed date. | DeepSeek | — | — | — | Official API · Hosted providers — pricing pending weekly sweep |
| GLM-5.3 Released after GLM-5.2; license not yet confirmed against GLM-5.2's MIT terms — verify before citing. | Z.ai (Zhipu AI) | — | — | — | Official API · Hosted providers — pricing pending weekly sweep |
| GLM-5.2 | Z.ai (Zhipu AI) | MIT | — | — | Self-host · Official API · Hosted providers — pricing pending weekly sweep |
| Qwen3.6-27B | Alibaba (Qwen team) | Apache 2.0 | 27B | — | Self-host (single high-memory GPU) · Official API · Hosted providers — pricing pending weekly sweep |
| Qwen3.6-35B-A3B A3B = 3B active parameters at inference (mixture-of-experts). | Alibaba (Qwen team) | Apache 2.0 | 35B / 3B | — | Self-host · Official API · Hosted providers — pricing pending weekly sweep |
| Qwen3 Coder 480B-A35B Coding-specialized mixture-of-experts release. | Alibaba (Qwen team) | Apache 2.0 | 480B / 35B | — | Self-host (large-scale infra required) · Official API · Hosted providers — pricing pending weekly sweep |
| Llama 4 Maverick License applies use-based restrictions above a 700M MAU threshold and carves out multimodal use in the EU — read the license before deploying. | Meta | Llama 4 Community License (EU multimodal carve-out) | — | — | Self-host · Hosted providers — pricing pending weekly sweep |
| Llama 4 Scout Same license family as Maverick — 700M MAU threshold and an EU multimodal carve-out apply. | Meta | Llama 4 Community License (EU multimodal carve-out) | — | — | Self-host · Hosted providers — pricing pending weekly sweep |
| gpt-oss-120b | OpenAI | Apache 2.0 | 120B | — | Self-host (large-scale infra required) · Hosted providers — pricing pending weekly sweep |
| Hermes 4 70B Nous ships Hermes 4 in multiple sizes; only the 70B variant is confirmed in our source set as of seed date. | Nous Research | — | 70B | — | Self-host · Hosted providers — pricing pending weekly sweep |
Per-provider pricing columns (Morpheus among them, not featured) populate as the weekly price sweep runs — see the Inference Price Index.
About this dataset
Seed data compiled from DeAI’s launch-sprint sourcing (last updated 2026-08-24). Publisher and license are stated only where well documented; parameter counts, context windows, and release dates we could not confirm are left null rather than estimated. Model names ending in a size suffix (e.g. “27B”, “480B-A35B”) reflect the publisher’s own naming, not an independent benchmark. This tracker is a Desk C data asset; see methodologyfor how it’s maintained.