Molmo 72B API: no seller prices the 72B tier, so here is the dated census, the token facts, and the sibling lane that is served

There is no priced Molmo 72B API to buy. Checked on 28 October 2026, no marketplace or host publishes a rate for the 72B model of the 2024 Molmo family. The prices that surface in search belong to two other tiers. One is the 7B port on a single host. The other is the newer Molmo 2 base models. This page holds the dated census behind that gap. It gives the 72B facts from the official card and config. It works the token math that the tokens queries ask for. It covers what running the open weights yourself takes. And it names the one lane of this family tree that this platform does serve, molmo2-er, with its live rate.

Last verified: 2026-10-28

What Molmo 72B is, and what it is not

Molmo 72B is the largest of the four models the Allen Institute for AI released on 25 September 2024, in the first Molmo family. The release post names them: MolmoE-1B, Molmo-7B-O, Molmo-7B-D, and Molmo-72B, the best of the four. The family was trained on PixMo, a set of under one million hand-collected image and text pairs. Its signature skill is pointing: the models can indicate a spot in an image, not only describe it. The paper makes the openness claim plain. No proprietary vision model distilled the data.

The 72B has one public checkpoint today, allenai/Molmo-72B-0924. Its model card states the build: a Qwen2-72B language backbone with an OpenAI CLIP vision encoder in front. The checkpoint carries the Apache-2.0 license. Its stated intent is research and educational use. The underlying Qwen2-72B base carries its own Tongyi Qianwen license, which the card links directly.

Three names trip up every Molmo search, and this page will repeat the map because the rest of the market never does. Molmo 72B is the 2024 vision-language heavyweight on this page. Molmo2-ER is the 2026 embodied-reasoning model built on a 4 billion base, and it is the lane this platform serves. MolmoAct2 is the 2026 action model that uses ER as its eyes. The publisher's own landing page now headlines the Molmo 2 generation. The 2024 flagship is no longer the marketed face of the family. That is the precedent in this page's title: one family tree, one lane served here, one big sibling with no seller anywhere.

On the publisher's own tests the 72B was the strongest open model of its release, and it ran against the closed leaders of 2024. These are the publisher's numbers, quoted as such from the card.

The one-line read: highest academic average of the set, second on human preference, 2 Elo points behind GPT-4o. The abstract says the same in one sentence. The 72B outperforms Claude 3.5 Sonnet and both Gemini 1.5 Pro and Flash, second only to GPT-4o.

Who sells Molmo 72B today: the census

People search "molmo 72b api", "molmo 72b price", and "molmo 72b tokens". Autocomplete on the family is live, checked today. The api string returns its malmo typo variant. The bare model string returns the "molmo 7b 72b" pair string. The publisher-qualified string returns "allenai molmo 72b 0924". Every one of those searchers wants a rate row. Here is what each place that could sell the model actually holds, read on 28 October 2026.

Three notes on the table. First, every row is dated, because host states move and this is the part of any page that goes stale first. Second, the absence is not a license wall in the simple sense. The checkpoint is Apache-2.0 and anyone may serve it. The card still states a research and educational intent. The Qwen2 base carries its own license terms too. A commercial seller owes both a read. What is missing is a seller who has done the serving work and published a rate. Third, three engines, fal, Nebius, and Hyperbolic, keep their model lists behind a login or a JavaScript-only surface. They returned no molmo evidence on the surfaces that are public, and the claim here stays scoped to the rows above.

Why every price you meet belongs to another Molmo

Search the money queries and the results sell you a different model. On the live result pages read today, the top hits for "molmo 72b price" and "molmo 72b tokens" were seven pages about the 7B port. Every one of them. The autocomplete string "molmo 7b 72b" exists because searchers themselves mix the tiers. Three price shapes circulate, and none of them is a 72B rate.

The first shape is the priced 7B port above, about 0.042 per run. A real listing, for a model nine times smaller. The second is the stale record class: a 0.2 per million token figure from a base Molmo 2 page that no longer sells. Directories repeat the figure without dates. The Molmo2-ER pricing guide decodes that record in full, because the ER variant inherits the same pollution. The third is the template page class, and it is new enough that nobody warns you about it. Directory sites generate a works-with page per model id, hundreds of them. One of them claims the 72B with a 128K context figure. The official config file for this exact checkpoint says 4,096. A 32 times inflation, printed as fact, on the first page of results.

That is the honest answer to the price query. The 72B rate does not exist. The pages that rank for the query are selling you the sibling or a template.

The tokens question, answered from the config

The "molmo 72b tokens" searches want hard numbers, and the official files give exact ones. The checkpoint's config sets the context window at 4,096 positions. That number decides the workload math more than any benchmark does.

Image tokens come from the CLIP side, and the checkpoint's own files give two numbers that are easy to mix up. The official preprocessor config sets a 336 by 336 base input with a 14 by 14 patch size. That is a 24 by 24 grid, 576 vision patches. The same file's image token length fields say 12 by 12, and the modeling code pools the patch grid 2 by 2 down to that size. So one base-resolution image costs the context window 144 image tokens, plus 14 framing markers, one after each of the 12 patch rows plus a start and an end. Both numbers come from the same official file, and the per-image figure is math on that config, not a measurement. Multi-crop mode raises it: the config allows up to 12 crops per image on top of the base view, each adding its own 144 token block, so one heavily cropped image can approach half the window by itself.

A perception loop that grounds every step against the same camera scene will refill the window on each call. This checkpoint has no published caching lever. Compare the served sibling before you budget. The molmo2-er row enforces an 8,192 token window, double the 72B, and bills a repeated input prefix at a cached rate. The Molmo2-ER pricing guide works that math with real rates. For the conversion between token counts, per-call meters, and per-hour figures, read the pricing units guide.

What running the 72B yourself takes

With no rate to buy, the open weights are the price you can act on, and the 72B makes you pay in gigabytes. The shipped checkpoint stores 32 bit weights, 293.2 GB across 83 files, per the repo's own storage record read today. A 16 bit load lands near half of that by math, about 146 GB. The card notes that 16 bit can change outputs against the 32 bit eval setting. A 4 bit quantization of 72 billion parameters is about 36 GB by the same math. That is half a byte per parameter, before any overhead. Community quantization pages for this model agree on that class of size.

The official repository is blunt about scale. Evaluating the 72B with its codebase takes multiple nodes, with a memory flag the readme names. For serving, the card supports vLLM with a cap, version 0.7.2 or older, until a preprocessing bug is fixed. The linked bug report closed on 11 October 2025, per the issue tracker read today. The cap note and the fix state disagree. Check current vLLM support before you bet a serving stack on either.

Two practical notes from the card's own FAQ, because they cost an afternoon if you meet them cold. Molmo errors on images that are not RGB, and it struggles with transparent images. Convert and flatten backgrounds first. For the smaller Molmo 2 generation, the local-run path has its own guide: Molmo 2 local inference usage. For card rental rates and the break-even math against metered APIs, the RTX 3090 account-rate guide and the RTX 3060 budget guide carry the measured tables.

The precedent: the served sibling lane

Here is what the family tree looks like from this platform. The 2026 embodied-reasoning lane is served. The molmo2-er model id is listed in the platform's model data and serves on the chat route. It has been listed since 6 October 2026. Metering is per million tokens. The served context window is 8,192 tokens. The 2024 heavyweight lane next to it has no row here either, and this page promises nothing about one.

That fence is the working precedent. Same publisher, same shop window. One lane is listed and metered at a published rate. The 72B lane has no row anywhere on the market, and none here. Spend runs against a prepaid wallet, and the wallet balance is the hard stop. A refused request bills nothing. The argument for that shape under unattended loops is the API bill shock guide.

The rest of the family has its own pages. The action model's written record is the MolmoAct2 pricing guide. The wider question of which big hosts serve robot and vision models at all has a dedicated census: do the big hosts serve robot models. And the register for a model with no rate card anywhere, worked to the last call count, is the OpenVLA-7B pricing guide.

What to do now

1. Read the fence, not the search results. The live rate for the served lane always renders on the pricing page. Every other number in this market belongs to another tier or a stale record.
2. Match the window to the job before the model. The 72B gives you 4,096 positions and 144 image tokens per base-resolution image. The served molmo2-er row gives 8,192 and a cached prefix rate. The workload decides, and the model page carries the full contract.
3. Write the client before any spend. The chat route is OpenAI compatible and the request shapes are documented today. Start at the quickstart and the OpenAI-compatible guide.
4. Re-check the census dates before you cite a host. If a 72B seller appears, the census table above is the part that goes stale first. You can sign up and create a key without talking to anyone. The prepaid balance is the spend stop.

Frequently asked questions

Is there a Molmo 72B API?

No priced one, as of 28 October 2026. The Hugging Face provider mapping on the checkpoint is empty. The router, DeepInfra, Novita, and Fireworks list no molmo row in the 72B's size class. Fireworks and Replicate price only other tiers of the family. One aggregator spec page prints, in its own words, that no API pricing data is available for this model. One template page claims a rate and fails its own spec check against the official config. The publisher ships weights and code with no product.

What is the context window of Molmo 72B?

4,096 positions, set in the checkpoint's official config file. The number is exact, not a rumor: the config's max_position_embeddings field says 4096. Aggregator pages print 8,192 or even 128K for this model. Both are wrong against the primary source. The 128K figure comes from a generated template page. Each image takes 144 tokens at base resolution, so the window fills after roughly 26 to 28 bare images with no text.

Is Molmo 72B the same model as Molmo2-ER?

No. They are two generations of one family tree with different jobs. Molmo 72B is the 2024 vision-language flagship on a Qwen2-72B backbone, a general image-understanding and pointing model. Molmo2-ER is the 2026 embodied-reasoning perception model on a 4 billion base, built for robot-facing scene work. It is the lane this platform serves on its chat route. The Molmo2-ER pricing guide carries that lane's full story.

What does Molmo 72B cost to run?

No per-token or per-call rate exists to quote, so the honest price is the self-host floor. The shipped 32 bit checkpoint is 293.2 GB. A 16 bit conversion lands near 146 GB, half the shipped size. A 4 bit quantization lands near 36 GB, half a byte per parameter. The official evaluation path for this model spans multiple nodes. Card rental and engineering time sit on top. For a metered alternative in the same family, the fence above carries the served lane's live rate.

Can I run Molmo 72B myself?

Yes, and the license is the easy part. The checkpoint is Apache-2.0 with a stated research and educational intent. The Qwen2-72B base carries its own license, which the card links. The official repository carries the code. The demo runs the smaller 7B-D. The model page lists llama.cpp, Ollama, and LM Studio quantization paths. Feed it RGB images with flattened backgrounds. Serve it on vLLM only after checking the version note on the card against the current bug state.

The short version: no seller prices the 72B anywhere. The price results in search belong to the 7B and the Molmo 2 tiers. The real 4,096 token spec sits in the official config. The family's served lane here is molmo2-er, with its live rate on the pricing page. Sign up to create a key and budget metered calls.

Last verified: 2026-10-28