MiniMax hints at M3 on SambaNova, provides no benchmark result
An undated MiniMax post mentions M3 and SambaNova hardware, while the companies' disclosed demonstration identifies M2.7 and provides no M3 setup or result.
By RuntimeWire Staff · Published
Primary source: MiniMax
Why it matters
MiniMax M3 is an approximately 428-billion-parameter open-weight model with approximately 23 billion activated parameters, built for long-context coding and agentic work. SambaNova has published strong M2.7 throughput figures, but no verified M3 configuration or result. Reproducible speed, cost and power measurements would matter for MiniMax's distribution and SambaNova's case for the RDU architecture behind its $11B valuation.
Yan Junjie's MiniMax posted an undated question asking how fast M3 was "about to run" on SambaNova hardware. The linked SambaNova post provides no readable technical details in the supplied material.
The available material does not establish whether this is a commercial deployment, a benchmark demonstration or a simple teaser. It provides no M3 tokens-per-second result, SambaNova system configuration, workload or availability information. SambaNova's public account of the companies' hardware demonstration identifies the model as M2.7, rather than M3.
Yan founded MiniMax in early 2022 after spending more than six years at SenseTime, where he became a vice president and vice-head of its research institute. MiniMax's investor materials identify him as chairman, CEO and CTO. He studied mathematics at Southeast University, earned an AI doctorate from the Chinese Academy of Sciences and conducted postdoctoral research at Tsinghua University.
M3 reflects Yan's effort to make long-running coding and agent workloads practical without abandoning model scale. If SambaNova implements M3, the work could test that architecture outside MiniMax's own serving environment and give another hardware vendor an incentive to distribute the model. The posts do not establish that such an implementation has occurred.
A model built around the cost of long context
MiniMax released M3 on June 1st, 2026, describing it as a native multimodal model for coding, computer use and agentic workflows with a 1-million-token context window. The MiniMax M3 model card states that the open-weight model has approximately 428 billion total parameters and approximately 23 billion activated parameters.
The central engineering claim is MiniMax Sparse Attention, or MSA. MiniMax says in its M3 announcement that MSA cuts per-token compute at a 1-million-token context to one-twentieth of the previous model's requirement, while delivering a 9x prefill speedup and a decoding improvement above 15x. Those are MiniMax's measurements, comparing M3 with its prior architecture rather than with competing accelerators or hosted services.
MiniMax has since released M3's weights, giving infrastructure providers the ability to optimize serving around their own chips and software. The model's size and long-context design would make a documented SambaNova run more consequential than a short chatbot demo. Coding agents can generate long sequences of tool calls, test output and revisions, forcing an inference system to keep moving data and producing tokens over an extended session.
That workload is also where headline throughput figures become easy to misread. Prefill speed measures how quickly a system processes the initial prompt and context. Decode speed measures token generation after that context has been loaded. An end-to-end result must account for both, along with batch size, input length, output length, precision and the number of chips involved.
SambaNova's verified MiniMax result uses M2.7
According to SambaNova's public account of its MiniMax M2.7 demonstration, the heterogeneous system used four Nvidia H200 GPUs for prefill and 16 SambaNova SN50 reconfigurable dataflow unit chips for decode. SambaNova said the system reached up to 850 tokens per second on short-context workloads and more than 450 tokens per second on long-context workloads, citing benchmarking by Artificial Analysis.
Those M2.7 figures provide a possible reference point for M3. The undated teaser does not say whether SambaNova has run M3, whether it would use the same configuration, or whether any eventual number would cover decode, prefill or a complete request.
SambaNova co-founder and CEO Rodrigo Liang spent more than two decades working on processors at Hewlett-Packard, Sun Microsystems and Oracle before starting SambaNova in 2017. His co-founders include Stanford professor Kunle Olukotun, an early architect of chip multiprocessor designs, and Stanford computer scientist Chris Re. Their wager is that AI inference benefits from hardware designed around data movement rather than forcing every stage of a model request through the same accelerator.
That wager is now heavily financed. In a press release, SambaNova said it completed the $1B first close of a Series F led by General Atlantic at an $11B post-money valuation. The company said the round included "significant investment" from Seligman Ventures and T. Rowe Price Associates. It listed Capital Group, BlackRock-managed funds, Intel Capital, Qatar Investment Authority and Vista Equity Partners among the other new and existing investors.
The funding gives Liang room to build chips and rack-scale systems. It also raises the standard for proof. SambaNova needs benchmark results that show how its RDUs compare on latency, throughput and cost per token under workloads customers will recognize.
The teaser leaves the important questions open
MiniMax's post invokes M3 without supplying evidence of a completed or planned benchmark. A credible result would need the hardware count, model precision, serving software, context lengths, batch sizes and separate prefill and decode measurements. Cost and power consumption would make the comparison useful for buyers rather than merely impressive in a feed.
Commercial availability remains another open question. The posts do not establish whether customers can use an M3 implementation on SambaNova hardware, or whether such an implementation exists.
MiniMax has been widening its infrastructure footprint alongside its model releases. RuntimeWire reported this week that RadixArk released Miles v0.1 to connect AI training and inference, another project built around making MiniMax models easier to operate across the computing stack.
Yan's M3 architecture and Liang's RDU system address different parts of the same constraint. MiniMax is reducing the work a long-context model needs to perform, while SambaNova is trying to move inference workloads through specialized hardware faster. A published M3 result could show whether those designs work well together, provided it includes enough detail for other engineers to reproduce it.