Hold data and its order constant across four sizes, and the only thing separating the dense 32B from the sparse 36B-A4B is architecture and parameter count [11]. That is the condition under which IFM's efficiency claim can be argued at all. MoVA pushes expert routing into multi-head attention rather than only the feed-forward layers, and IFM says it stays compatible with FlashAttention, grouped-query attention and sparse attention [6]. IFM reports the 36B-A4B landing only slightly below its dense 32B sibling under the same training conditions while activating fewer parameters per token [7]. The arithmetic behind "fewer": about 4 billion of 36 billion, roughly 11 percent [5][1], against a 32B model that activates all of itself, so about eight times fewer activated parameters per token [3]. "Slightly below" is not a number and the release does not supply one [7], so the published evaluation results are where that argument gets settled [3].
The Uno adapters carry the same caveat with less cover. IFM reports roughly threefold generation speedup on the 0.9B and 7B models with no quality loss, from its own evaluations, unaudited [8]. Speedup is a property of a serving configuration, not of a model: batch size, context length, quantization, concurrency, hardware. The release reports none of them [8]. For 3x to land on your stack, your decode shape has to match whatever IFM measured, and right now you cannot check that.
Then the token counts. One passage says approximately 20 trillion per model, about half of it synthetic [9][6]. Another says 22 trillion shared across the four mid-sizes [11], and IFM does not reconcile the two [12]. Two trillion tokens is a 10 percent gap [4]. Run the reasoning-trajectory share through both bases and nearly 17 percent is either 3.4 trillion tokens or 3.74 trillion, a spread of 340 billion [10][5]. That is exactly the class of question a training log and a dataset card answer and a model card does not [3], which is the whole argument for shipping them.
The synthetic half deserves its own line. Synthetic tokens have no third-party licence holder, so they fall cleanly on the publishable side of "data where licences permit" [3], and the reproducibility question moves to the generator and prompt set that produced ten trillion of them [9].
Adoption cost is where the openness thesis meets a budget. Nobody outside a handful of labs is re-running a 375B-A23B pretrain that activates about 23 billion parameters per token, roughly 6 percent of the total [5][2]. The test an ordinary buyer can afford is rebuilding 0.9B or 3.7B from the recipe, then checking the released checkpoints and losses of the large runs for internal consistency [3][18]. Read the individual cards while you are in there: the 32B FP8 card states a native 524,288-token context window from midtraining onward [14], which is 2 to the 19th [7], so at least that figure came from an engineer rather than a round-up.
IFM was launched by MBZUAI in May 2025 and also runs research hubs in Silicon Valley and Paris [16]. Its release language moves between present and future tense, including a promise that the full agentic post-training code base will be released [15]. The artifacts already up are enough to work with [13]; the ones written in future tense are the ones a procurement clause should name, model by model [18].
