Hardcore Reviews
Hardcore Reviews

Four Open Models, One 16B Base: Why Four Paths?

NVIDIA's four Cosmos 3 Nano checkpoints are called four new open-weight checkpoints in the official group blog. All four share one 16B base, the Mixture-of-Transformers (MoT) backbone and the OpenMDW1.1 license, so parameter count, license and backbone are identical cell by cell; the differences lie in input contracts, generated output and who holds physical ground truth. This comparison follows four post-training paths, Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual: from text prompts plus World Scenario Map controls to synchronized multi-view camera video and LiDAR, from the causally distilled streaming variant, through Depth, Segmentation, Edges and Blur spatial controls rendered into color observations, to video predicted from actions. Post-training is described only in the model cards' two stages, post-training then causal distillation, not as four separate recipes. All four cards list Release Date Hugging Face 10/20/2026, kept alongside the NVIDIA blog and the Hugging Face group blog as three date frames rather than merged into one launch day. No leaderboard scores, no repository metadata registry and no single-product launch report.

Published November 9, 20269 min read
<!-- cosmos3-nano-four-paths-comparison-review | review | Four Open Models, One 16B Base: Why Four Paths? -->

Put NVIDIA's four Cosmos 3 Nano checkpoints side by side: the question is not what each generates, but why one 16B base became four post-training paths. The official group blog calls them four new open-weight checkpoints, all post-trained examples of Cosmos 3 Nano, each a post-trained variant of a shared base. Open-weight means weights available for further post-training, not that training data, the full recipe and every engineering detail are public; the title's loose label should not be read that way. So why split one 16B base into Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer and Cosmos3-Nano-Sim-Bimanual? They differ in input contracts, output and who holds physical ground truth, not in parameter count, license or backbone, which are identical.

Place this piece within the batch first. It covers the mechanics of four paths from one base and their input-output contracts. The hotspot column covers the GIF and SVG additions in OpenAI Images 2.6, the open-source column digs into oxc-project/tsrs, and the upgrade SOP walks through Google ADK 2.14.0 breaking changes; each tracks one object. The GitHub tools metadata comparison from November 7, 2026 is a snapshot table of API metadata for five unrelated repositories at one capture point: language, license, star count, creation date and runtime prerequisites. The world-model single-product feature reports one launch. This piece does neither. It is a problem-driven teardown: shared fields are identical, so the difference does not live in stars or creation dates, and a snapshot table cannot carry it.

Set the hook with official wording only. All four model cards list 10/20/2026 as their Hugging Face Release Date. The NVIDIA official blog then announced new Cosmos 3 Nano simulation variants, and the official Hugging Face blog documented the four checkpoints together for interactive world simulation.

What the four share

Back to the official wording. The group blog states All four models are post-trained examples of Cosmos 3 Nano and have 16B trainable parameters. The models use OpenMDW1.1, permitting commercial and non-commercial use under its terms. Cards repeat Network Architecture: Mixture-of-Transformers (MoT), Number of trainable model parameters: 16B, This model is released under OpenMDW1.1. All four are post-trained variants of the Cosmos3-Nano base with 16B trainable parameters, the Mixture-of-Transformers (MoT) backbone and OpenMDW1.1: four branches off one starting point, not four from-scratch trainings. The cards also share one field structure:

text
Model Overview
Description
License
Deployment Geography
Use Case
Release Date
Model Architecture
Inputs
Outputs

A shared schema puts all four inputs and outputs into the same named slots. Post-training on multi-sensor data yields Transfer-Auto, then causal distillation yields the streaming variant; the Sim-Bimanual card reads post-training ... followed by causal distillation. Two stages, not one recipe per model. On the license, only official wording is used, and a clause-by-clause teardown already exists in the open-weight license comparison from October 23, 2026; here it is one shared fact, with no clause comparison and no legal opinion.

The table registers functional-contract text only. Parameter count, license and backbone are identical, so they sit in one shared-fact line outside it; numbers of different units stay inside the section for the model they belong to. Unstated capabilities are recorded as "the official pages do not state it".

Shared fieldCosmos3-Nano-Transfer-AutoCosmos3-Nano-Sim-AutoCosmos3-Nano-Sim-TransferCosmos3-Nano-Sim-Bimanual
Official one-line roleGenerate multi-view driving data with varied weather, lighting, and time of day.Stream synchronized camera and LiDAR observations for driving simulation.Turn simulator controls and a text prompt into photorealistic video.Predict video from robot or camera actions.
Input conditionsText prompt, World Scenario Map (WSM) controls per view, aligned LiDAR controls when LiDAR is requested.Per-camera captions, synchronized WSM controls, optional initial RGB images, HD-Map range controls for LiDAR.Spatial controls (Depth, Segmentation, Edges, Blur) and a text prompt.Initial image plus control inputs, injectable at any point in a rollout.
Generated outputSynchronized multi-view camera video and LiDAR.Successive small chunks of synchronized camera video and LiDAR.Color observations a policy can consume.Video in small latent-frame chunks, MP4 or streamed frames.
Use caseSynthetic multi-sensor driving data for training and evaluating perception and driving models.Fast streaming and long, interactive rollouts.Simulator keeps physics and ground truth; the model renders.Robotic policy evaluation, reinforcement learning training, synthetic demonstration data.
Path linkSource checkpoint for causal distillation into Sim-Auto.Distilled from Transfer-Auto; adapt the source first for a new rig or domain.Parallel to Sim-Bimanual: controls to observations.Parallel to Sim-Transfer: actions to video.

The identical cells are not repeated in the table: all four are post-trained variants of Cosmos 3 Nano with 16B trainable parameters, OpenMDW1.1 and the Mixture-of-Transformers (MoT) backbone; the official model card lists the same field structure.

Transfer-Auto: an offline, bidirectional multi-view data generator

Transfer-Auto's role is Generate multi-view driving data with varied weather, lighting, and time of day. The card calls it a bidirectional model for offline synthetic driving data generation: bidirectional and offline, seeing a whole segment at once to make aligned multi-view, multi-sensor driving data in batches. Input, verbatim: Generation is conditioned on a text prompt, World Scenario Map (WSM) controls for each camera view, and aligned LiDAR controls when LiDAR output is requested. The prompt governs appearance, WSM controls govern structure and traffic motion, and LiDAR adds an aligned channel.

Output, verbatim: generates synchronized video across up to 11 camera views, including fisheyes, with up to 11 camera views including fisheyes and camera video and LiDAR generated jointly and aligned. It is also, verbatim, serves as the source checkpoint for causal distillation into Cosmos3-Nano-Sim-Auto: the batch generator is the checkpoint that gets distilled, not the end of the line.

Sim-Auto: the causally distilled streaming variant

Sim-Auto's role is Stream synchronized camera and LiDAR observations for driving simulation. The official wording fixes the relation: the causally distilled streaming variant of Transfer-Auto, with the card labeling it a post-trained, causal-distilled variant. Generation, verbatim: generates successive small chunks of synchronized camera video and LiDAR in just a few denoising steps. A whole-segment bidirectional pass becomes few-step denoising emitting successive small chunks, continued from generated history and refreshed controls. Those chunks and steps are this model's own mechanism, not numbers set against others.

Input, verbatim: Provide per-camera captions and synchronized WSM controls, optionally adding initial RGB images. For LiDAR generation, include HD-Map range controls aligned with the camera controls. Closed loop needs a simulation runtime and WSM renderer to refresh controls from driving inputs, and for a new rig or domain the official order is For a new sensor rig or domain, adapt Transfer-Auto first, then distill it into Sim-Auto.

Sim-Transfer: a renderer that leaves physics with the simulator

Sim-Transfer's role is Turn simulator controls and a text prompt into photorealistic video. It cuts along a different axis: who owns physics and who owns appearance. Verbatim: Your simulator computes motion, contacts, and physical state. Sim-Transfer uses its spatial controls and a text prompt to render color observations that a policy can consume. The simulator computes physics; the model renders. Its spatial controls are listed by name: Depth: relative distance from the camera. Segmentation: regions and object boundaries. Edges: scene contours. Blur: coarse appearance and layout. Four enumerated channels, not comparable numbers. A following line reads Text guides appearance; physics and ground-truth state stay with the simulator.

Its relation to Sim-Bimanual is a parallel, verbatim: Sim-Bimanual predicts visual consequences from actions, while Sim-Transfer turns a physics simulator's controls into generated observations. As for whether this one streams, the official pages do not state it, and this piece does not fill it in.

Sim-Bimanual: a different embodiment, from actions to video

Sim-Bimanual's role is Predict video from robot or camera actions. The official wording says it simulates how an environment evolves in response to a bimanual robot's actions, and the card labels it a post-trained, causal-distilled variant of the Cosmos3-Nano base model: still on the same 16B base, now action-driven rather than multi-sensor driving. Input and generation, verbatim: Given an initial image and a sequence of control inputs, the model generates video in small latent-frame chunks conditioned on cached history and incoming robot actions or camera trajectories. Video comes in small latent-frame chunks conditioned on cached history and incoming actions; control inputs can be injected at any point in a rollout, and Supported embodiment reads AgiBot.

Output belongs with one official qualifier that must travel with it: the card reads Video: MP4 or streamed video frames and Generates video at 480p with an output frame rate of 30 frames per second (fps). Output video frame rates describe playback timing; inference throughput is reported separately. The 480p and 30 fps describe playback timing, and the official line separates that from inference throughput, so both stay in this section, are not set beside other models' quantities and are not turned into a speed conclusion.

How the four paths fit together

Offline versus streaming is the only chain where the official pages state an upstream and downstream relation: The adapted model can be used directly for offline data generation or undergo causal training and few-step distillation for fast streaming simulation. That is the development path used to create Cosmos3-Nano-Sim-Auto. The card writes two stages, derived from Cosmos3-Nano in two stages: first, post-training ... to produce Cosmos3-Nano-Transfer-Auto; then, causal distillation, and the Sim-Bimanual card reads post-training ... followed by causal distillation. The second axis is who owns physics, with Sim-Transfer leaving it to the simulator; the third is embodiment, with Transfer-Auto and Sim-Auto organized around the World Scenario Map and WSM controls, Sim-Bimanual on AgiBot, and Sim-Transfer on rendering. All four share 16B, OpenMDW1.1 and MoT.

Runtime preconditions and the boundary NVIDIA draws itself

Runtime preconditions are only the two the official wording names: Sim-Auto needs a simulation runtime and WSM renderer to refresh controls from driving inputs, and Sim-Transfer needs a simulator to compute motion, contacts, and physical state first. Where no requirement is stated, none is filled in, and no hardware specification is listed. More important is the boundary NVIDIA draws itself, verbatim: These problems remain hard: the work ahead is to measure where the models faithfully support policy learning and evaluation, understand their limits, and improve where they fall short. That puts how faithfully the four support policy learning and evaluation in the category of work still to be measured, not a delivered conclusion. This piece records the four only as a starting point, with boundaries referred back to per-model measurement; second-hand summaries are not promoted to official conclusions.

Scope notes: three dates and what is left out

Three official date readings sit side by side and are not reconciled. The Release Date field of all four cards reads Release Date - Hugging Face 10/20/2026. The NVIDIA official blog page itself is dated October 21, 2026. The official product page lists the same entry as [October 22, 2026]. Each is quoted as written, with no single launch date stitched together. A separate claim dates the official Hugging Face group blog to one particular day, but it exists only in second-hand roundups with no verbatim official wording, so this piece does not print that day or treat a roundup date as an online date.

Whether a formal technical report exists is inconsistent across two official pages, so it is not raised; a stage count larger than two appears only in second-hand summaries, so the two-stage path is written as the cards give it. Unconfirmed embodiment names, control-value specifications and cross-model arithmetic do not enter the text. Numbers the official wording does not state are neither filled in nor guessed at, and numbers of different units stay in their own model's section.

FAQ

Q1: Are these four independently trained models?

A1: No. All four are post-trained examples of Cosmos 3 Nano, variants of one 16B base with 16B trainable parameters, the Mixture-of-Transformers (MoT) backbone and OpenMDW1.1. Those fields are identical, so differences lie in input contracts, outputs and use cases.

Q2: Why split one 16B base into four paths?

A2: The cuts are orthogonal axes: Transfer-Auto is offline bidirectional and distills into the streaming Sim-Auto; Sim-Transfer leaves physics with the simulator; Sim-Bimanual switches embodiment to actions-to-video. All four share 16B, OpenMDW1.1 and MoT, so differences lie in input, output and physical-ground-truth ownership.

Q3: What fields is the comparison table organized around?

A3: The functional-contract fields the four cards share: one-line role, input conditions, generated output, use case and path link, plus common ground. Fields are uniform, only contract text is compared, and numbers of different units never share a column. Unstated capabilities are recorded as "the official pages do not state it".

Q4: Why are there several dates, and which one should I use?

A4: Three official readings sit side by side, unreconciled: the cards read Release Date - Hugging Face 10/20/2026, the NVIDIA official blog page itself is dated October 21, 2026, and the official product page lists the same entry as [October 22, 2026]. Each is quoted as written, with no single launch date stitched together.

Q5: How should the 480p and 30 fps of Sim-Bimanual be read?

A5: Those two numbers appear only inside the Sim-Bimanual section and describe output playback timing. The official line immediately states Output video frame rates describe playback timing; inference throughput is reported separately., separating playback frame rate from inference throughput. They are not a speed conclusion, and are neither set beside other models' quantities nor converted. The card's output reads Video: MP4 or streamed video frames, and Supported embodiment reads AgiBot.

Related reading

This article is AI-assisted and human-edited. Last updated: 2026-11-09

FAQ

Are these four independently trained models?
No. All four are `post-trained examples of Cosmos 3 Nano`, variants of one 16B base with `16B trainable parameters`, the `Mixture-of-Transformers (MoT)` backbone and `OpenMDW1.1`. Those fields are identical, so differences lie in input contracts, outputs and use cases.
Why split one 16B base into four paths?
The cuts are orthogonal axes: Transfer-Auto is offline bidirectional and distills into the streaming Sim-Auto; Sim-Transfer leaves physics with the simulator; Sim-Bimanual switches embodiment to actions-to-video. All four share 16B, OpenMDW1.1 and MoT, so differences lie in input, output and physical-ground-truth ownership.
What fields is the comparison table organized around?
The functional-contract fields the four cards share: one-line role, input conditions, generated output, use case and path link, plus common ground. Fields are uniform, only contract text is compared, and numbers of different units never share a column. Unstated capabilities are recorded as "the official pages do not state it".
Why are there several dates, and which one should I use?
Three official readings sit side by side, unreconciled: the cards read `Release Date - Hugging Face 10/20/2026`, the NVIDIA official blog page itself is dated `October 21, 2026`, and the official product page lists the same entry as `[October 22, 2026]`. Each is quoted as written, with no single launch date stitched together.
How should the 480p and 30 fps of Sim-Bimanual be read?
Those two numbers appear only inside the Sim-Bimanual section and describe output playback timing. The official line immediately states `Output video frame rates describe playback timing; inference throughput is reported separately.`, separating playback frame rate from inference throughput. They are not a speed conclusion, and are neither set beside other models' quantities nor converted. The card's output reads `Video: MP4 or streamed video frames`, and Supported embodiment reads `AgiBot`.

Related

Hardcore Reviews

Open Weight, But How Open? Five Model Licenses Compared

Five open-weight models crowded into the same mid-October 2026 window, all saying 'open source' - and the substance spans three tiers. Tier A, genuinely open: Reflection Beam under Apache 2.0 with zero riders and ungated downloads; DeepSeek V4.1 Flash 1017 under MIT with zero riders; NVIDIA Nemotron 3.5 Bolt 74B-A9B under OpenMDW-1.1 and the only one of the five to publish weights, training data, and recipes together. Tier B, open weights with riders: Meta Muse Spark Open carries an attribution duty that extends to distilled products (models trained on its outputs must also display Built with Muse Spark Open), a 20-million-dollar annual-revenue threshold for MaaS resale, and gated HF downloads - legally not open source. Tier C, promise unfulfilled: Mistral Large 4 had no weights as of October 23, a license with no name, and an AA classification of proprietary. On the shared ruler of AA Intelligence Index v4.3.2: GLM-5.3 at 45, Kimi K3 at 44, DeepSeek 39.5 about even with Muse at 39, Mistral 38, and Beam 30 (Beam has no public API; its score comes from self-hosted inference). Self-reported DeepSWE and Terminal-Bench numbers use different harnesses and versions and cannot be compared directly. Hardware accessibility spans two orders of magnitude, from Bolt's single 24GB card to Muse's 700GB+ NVFP4 and Beam's 294GB checkpoint (about 400GB of memory to start). The piece includes an eight-dimension license table, a watering-down spectrum (attribution > revenue threshold > gated download > late weights > unnamed license > closed data), and per-use-case protocol advice; single-model deep dives for Muse and Beam live in their own articles.

Oct 23, 20268 min read
Hardcore Reviews

Closed API vs Open Weights: What Does One Image Really Cost

With ChatGPT Images 2.5 and Ant's open-source LLaDA-Image landing in the same week, text-to-image has split into closed APIs versus self-hosted open weights. This review ignores image quality and runs the cost-and-control numbers instead: five routes - closed APIs, self-hosted open weights, per-second third-party inference platforms, local consumer hardware, and domestic cloud APIs - with per-image cost projected at two volumes (100 and 10,000 images per day), plus a comparison table and scenario-based selection (hobby use, e-commerce batch, data-sensitive industries, brand-style fine-tuning, maximum quality). It flags four traps: undeclared licenses, cold starts on per-second billing, Chinese text rendering, and cross-border data transfer. Explicitly scoped apart from our 8-26 capability review of reasoning image models. Representative comparison, not hands-on benchmarking; pricing per official sites.

Sep 9, 20269 min read
Hardcore Reviews

TrnRT, Kindi, Baseten: One License Forbids Managed Service

Three inference engine repositories, TrnRT, Kindi and Baseten, on one comparison table, with license terms and hardware bets as the axis rather than a buying recommendation or benchmark ranking. Apache-2.0, Elastic License 2.0 and AGPL-3.0 coexist; one LICENSE says verbatim that the software may not be provided to third parties as a hosted or managed service, and GitHub reports NOASSERTION. Hardware claims from NKI megakernels on Trainium2 to GB300, B300, Quadro K420 and RTX 3080 are labelled as README self-reports. Stars are cumulative totals, with no growth score or single-day delta.

Nov 16, 20269 min read