Open models are making intelligence more available. They are not necessarily making power more distributed.
When the model becomes portable, where does power move? The Atlas tracks which families become infrastructure, how freely intelligence moves, who benefits from proliferation, and where dependency re-forms.
SCOPE Open-weight language, reasoning, multimodal-understanding and agentic model families. Image, video and audio generation families are outside this atlas.
Select one of the five lenses to change the claim, axes, highlighted families and strategic read.
Gravity = whether other builders are increasingly organising around a family as infrastructure.
Qwen derivatives — the clearest current substrate signal
How quickly attention, adoption or ecosystem activity is increasing now.
How much the ecosystem is building around this family.
The explosion in downloadable models is not producing an equally distributed ecosystem. A few families are becoming the bases that everyone else builds on.
A small set of families is attracting disproportionate downstream activity: derivatives, formats, provider support, tooling and developer habit. Qwen is the clearest current case, while Llama, Gemma, DeepSeek and Mistral show different forms of installed or deployment gravity.
Gravity is not model quality. It asks whether other builders increasingly organise around a family as reusable infrastructure.
Open weights can decentralise access while still reconcentrating ecosystem power around a handful of default substrates.
This read weakens if derivatives, provider support and tooling start dispersing across several families instead of compounding around the current anchors.
A monthly structural delta rather than a news feed. Only changes significant enough to alter the market read make it here: what changed, the mechanism underneath it, and the next signal that would strengthen or weaken the interpretation.
Derivative gravity continues to widen beyond a single flagship release.
More downstream models, formats and tooling make Qwen increasingly useful as a default substrate.
Watch whether ecosystem gravity survives changes in flagship licensing and keeps widening versus other base families.
AMD is no longer only optimising other open models; Instella makes it an original model publisher.
First-party models become proof points and optimisation targets for Instinct and ROCm.
If commercial licensing broadens, model publishing could become a more direct accelerator-distribution strategy.
A new from-scratch family entered the frontier cohort before building comparable provider or derivative depth.
Technical capability can now appear faster than distribution and ecosystem habit.
The test is whether provider support, quantisations and downstream builders arrive before the next frontier cycle.
LG and SK Telecom now represent distinct industrial routes into sovereign foundation models.
National capability is being funded through industrial and telecom balance sheets, not a single state champion.
Watch which programme converts sovereign funding into external developer, provider and enterprise adoption.
Its ecosystem remains large even as relative new-family momentum is weaker than Qwen and newer frontier entrants.
Installed tooling and deployment support decay more slowly than benchmark leadership.
The question is whether Meta renews release momentum or Llama becomes a durable legacy substrate.
The Arabic-first sovereign precedent remains important, but newer UAE programmes now carry more current strategic weight.
Sovereign model programmes can be superseded even when the institutional capability they created persists.
A new Jais generation with stronger capability or adoption would move it back into the active universe.
The families that matter for understanding where gravity, conversion, liquidity, complementary-asset power and dependency are moving.
Watch these for compounding gravity: derivatives, tooling, provider coverage and developer habit can make an open family powerful without making it closed.
These matter because frontier capability is no longer owned by a stable set of labs. The question is which challengers convert technical position into durable ecosystem pull.
Here the model may be subsidised because the patron monetises chips, cloud, devices, enterprise software or consumer distribution somewhere else in the stack.
These families are not best understood through global benchmark rank alone. Their economic return can be language coverage, national capability, institutional learning or reduced external dependency.
Coding, computer-use and multimodal specialists can win on task economics even if they never become general-purpose ecosystem anchors.
These families matter because portability and cost can outweigh frontier capability when models are placed on laptops, devices and edge systems.
As base weights commoditise, architecture, training recipes, alignment and reproducibility become separate sources of leverage.
Historical families and emerging candidates stay visible without being plotted as if they have equal current weight. Each entry has a specific promotion trigger.
Important Arabic-first sovereign precedent, but the current competitive UAE model story has moved toward newer programmes such as K2 Horizon and Falcon H1.
The official Jais family remains publicly available, but the tracked reference generation predates the current 2026 UAE frontier programmes and no new Jais flagship is captured in this audit.
A new Jais generation with materially stronger capability, release cadence or adoption.
Strategically important as an early data-platform-backed open model, but its flagship reference is from 2024 and no longer shapes the current frontier.
The official DBRX-Instruct checkpoint dates to April 2024. No current successor with comparable strategic visibility is captured in the September 2026 audit.
A new Databricks foundation-model release or a renewed ecosystem/adoption signal.
Retains downstream use, but the public family is no longer central to the 2026 capability or release-cadence race.
The current tracked official reference remains Yi-1.5 rather than a new 2026 generation, while newer Chinese families have moved the capability and distribution frontier forward.
A new flagship generation that restores capability, provider or derivative momentum.
An early Chinese open-weight contender with weaker current public momentum than Qwen, DeepSeek, GLM, Kimi or MiniMax.
The official Hugging Face organisation does not currently show a fresh 2026 flagship with adoption or provider evidence comparable to the active Chinese frontier families.
A current flagship release plus credible adoption or provider breadth.
Useful evidence of a data-cloud company publishing an open model, but the 2024 flagship is now primarily historical context.
The official Snowflake Arctic checkpoint dates to April 2024; no 2026 successor is captured in the current audit.
A new Snowflake open model or a renewed strategic model programme.
A strategically important European open foundation-model programme with active multilingual data, tokenizer and long-context model work, but not yet enough capability or adoption evidence for the active map.
The programme is actively publishing data/tokenizer work and an experimental 9B/256K SFT checkpoint. That checkpoint is derived from a Qwen3 9B base and is explicitly described as an experiment rather than a reasoning-specialised production flagship.
A production-grade flagship with clear adoption, independent evaluation or ecosystem gravity.
Apriel 1.6 is a credible compact multimodal reasoning family, but it has not yet built enough current ecosystem gravity to displace an active entry.
ServiceNow released Apriel-1.6-15B-Thinker in December 2025 as a 15B multimodal reasoning model with continual pretraining plus SFT/RL. The current audit does not show a comparably material 2026 adoption or release signal.
Sustained third-party adoption, derivatives or provider breadth beyond the initial release.
The 39B / 5B-active G9v3 release has unusually strong independent results for its size, but the family is too new to establish durable strategic significance.
The August 2026 Apache-2.0 preview is 39B total / 5B active with 131K context. Artificial Analysis currently places it well above the median for its size class, but only one API provider is tracked and the Hugging Face family is still small.
Repeat releases plus sustained ecosystem adoption or broader provider distribution.
A credible Apache-2.0 36B family with 512K context and reasoning/agentic positioning, but ByteDance is already represented on the active map by the more strategically differentiated UI-TARS line.
Seed-OSS-36B-Instruct is Apache-2.0, downloadable and configured for a 524K maximum position window, but its reference release dates to 2025 and no newer family-level signal currently justifies a second ByteDance active slot.
A major new Seed-OSS generation or ecosystem growth that makes the general family independently material.
Strategically relevant agentic post-training and infrastructure layer, but N1 releases are explicitly built on upstream foundations rather than forming an independent base-model family.
The public NEX-N1 repository lists DeepSeek-V3.1, Qwen3-32B, Qwen3-30B-A3B and InternLM3-8B NEX variants, making the upstream dependency explicit.
A from-scratch foundation model, or enough downstream ecosystem gravity to justify tracking post-training ecosystems as a separate primary class.
The curation record shows what entered the active universe, what is being watched, and what moved to the archive—plus the rationale for each decision.
Adds the missing AMD hardware-patron model story; trained from scratch on AMD Instinct hardware with an unusually open training recipe.
Current from-scratch frontier challenger with MIT weights and strong independent capability evidence.
Adds a second distinct Korean sovereign patron: a telecom-led 688B/33B-active foundation model trained from scratch.
Moved out of the active map: historically important Arabic sovereign family, but newer UAE programmes now carry more current strategic weight.
Active European foundation-model programme; waiting for a flagship with stronger capability/adoption evidence.
Strong efficient enterprise/multimodal candidate; ecosystem gravity is still too early.
Promising independent results for its size; too new to establish durable market significance.
Kept outside the base-family map as a strategically relevant post-training ecosystem built on upstream foundations.
Foundational historically; no longer a central current ecosystem choice.
Important instruction-tuning moment, now superseded by newer base families.
Influential post-training line whose role is now better represented by newer ecosystems.
Critical open-model infrastructure historically, but not a current competitive family.
No longer showing enough capability or ecosystem momentum for the active universe.
Removed from the active universe because current evidence is too weak to defend family-level strategic significance.
Tracks compounding ecosystem pull: derivatives, tooling, provider support, formats and developer habit.
Compares technical strength with downstream gravity to expose the gap between a strong model and a strategic winner.
Separates downloadable weights from practical legal, runtime, hardware and provider portability.
Looks for the complementary asset—silicon, cloud, devices, distribution, enterprise software or sovereign capability.
Surfaces upstream-base, patron, infrastructure, runtime and distribution dependencies that remain after weight access.
Tracks the monthly delta between curated snapshots, then states the mechanism and the next signal to watch.
Underlying inputs, not product destinations: capability, cost efficiency, momentum, gravity, liquidity and dependency. Scores are directional editorial indices, not measured market share. The diagnostic views use those inputs to answer structural questions; the dossier exposes the source-backed facts and caveats underneath them.
Viewer architecture: browsing, local model search, filtering, graph interactions, shareable URL state and model dossiers run locally after the initial load and do not trigger AI inference, remote search, database queries or Atlas backend requests. External source links leave the app.
Release identity, parameter scale, context, licensing, supported modalities and documented deployment paths.
Capability, price and speed where a current independent evaluator covers the exact model. Vendor benchmarks are never presented as independent measurements.
Hugging Face derivatives/downloads and OpenRouter routed usage are evidence of activity on those platforms—not global market share, revenue or quality.
Gravity, liquidity and dependency combine heterogeneous signals. Conversion and Power are diagnostic views built from those inputs, not separate measured facts. Every plotted score is directional synthesis; dossiers expose the evidence and caveats underneath.
Moving Parts follows the commercial and strategic shifts forming around emerging technologies—including the ideas behind this Atlas. Read the latest pieces, subscribe for new field notes and second-order reads, or come back for the next monthly cut.