MiniMax M3.1 Flash Is Now Served on Tenstorrent: What a 428B Model on Tensix Actually Proves

Tenstorrent's newsroom post is dated 22 October 2026, and it says something narrower and more interesting than "another cloud added another model": MiniMax M3.1 Flash, the 428-billion-parameter mixture-of-experts model MiniMax released on 19 October, is now available on Tenstorrent Cloud with both prefill and decode running on Blackhole Galaxy systems. No NVIDIA accelerator appears anywhere in that serving path. What Tenstorrent published is an architecture, a quantization scheme and a tuning methodology. What it did not publish is a single throughput number, a latency number, or a price for its own serving. That omission is the part that decides whether this is a landmark or a demo, and it is worth more attention than the model's benchmark table.

Tenstorrent is the chip company led by Jim Keller, and its pitch has always been that AI compute should not be a single-vendor stack. Until this announcement that pitch had a credibility problem at the top end: the largest open models were routinely served on NVIDIA hardware, with AMD as the alternative and everything else confined to small-model experiments. A 428B-parameter mixture-of-experts model running prefill and decode on Tensix silicon, connected by Ethernet rather than a proprietary fabric, is the first deployment that makes the pitch with a model of this size and this recency.

What Tenstorrent put on the record

The announcement is specific about the serving topology and vague about every number a buyer would use. The specifics first, because they are genuine engineering claims:

• Availability — Tenstorrent Cloud, at console.tenstorrent.com, through the Playground or connected to an agent harness.

• Serving path — both prefill and decode run on Tenstorrent Galaxy, specifically a supercluster of Blackhole Galaxies. The systems are split by workload: one pool handles prefill, another handles decode, and work moves between them over direct, low-latency Ethernet links.

• Memory — each Blackhole Galaxy carries 1 TB of DRAM plus several terabytes of SSD. Weights and the KV cache stay resident in DRAM; the SSD holds offloaded KV cache so a returning agent turn does not have to recompute its context.

• Quantization — MLP weights, attention weights, the indexer and the shared expert are stored in MXFP8; the routed experts are stored in NVFP4.

• Model shape — sparse attention replaced global grouped-query attention, and dense MLPs appear only in the first three layers.

• Tuning data — Tenstorrent says it tuned the deployment on SemiAnalysis's AgentX coding-agent traces, meaning real multi-turn sessions and subagent fan-out rather than short chat benchmarks.

• Price quoted — $0.15 per million input tokens, $0.60 per million output, $0.015 per million cached input. That is MiniMax's list price; the announcement does not publish a separate Tenstorrent rate, so it is not yet possible to say whether running the model here costs more or less than running it on GPUs.

The deployment exists because the weights came out

The single most important fact behind this announcement is not the silicon. It is that MiniMax published the weights at all. The 252.4 GB repository went up on Hugging Face on 23 October, and Tenstorrent's post is dated a day earlier — which suggests the port was built with partner access rather than from a public download. A third-party hardware vendor cannot port a model it cannot obtain, and Tenstorrent is not a MiniMax subsidiary serving its own checkpoint. This is the mechanism by which open weights create hardware competition: every lab that publishes a checkpoint turns every willing chip vendor into a potential distribution channel, and every chip vendor that supports the model makes the lab's weights more valuable in turn.

The same logic explains why the closed frontier models do not appear in this kind of announcement. A vendor can rent you an API endpoint for a proprietary model, but it cannot quantize the weights to its own numeric formats, tune its kernels against the model's sparse-attention pattern, or split prefill and decode across its own fabric. The port is the product, and the port requires the file.

That is also why the "open" in this story needs to be separated into two different things. Tenstorrent's software stack — TT-Metalium, TT-NN, TT-Forge — is open source, and that is what makes a third-party port tractable. MiniMax's weights are open weights, not open source. The repository is 252.4 GB across 48 safetensors shards plus a 2.3 GB speculative-decoding drafter, and it ships under the MiniMax Community Licence, which requires attribution on commercial deployments and prior written permission once annual revenue passes $20 million. Two different opens, one serving path.

Why the serving path is split in two

The prefill/decode split is not a Tenstorrent idiosyncrasy; it is the standard answer to a problem agent workloads create. Prefill is compute-bound and happens once per prompt. Decode is memory-bandwidth-bound and happens once per generated token, for as long as the answer runs. Put them on the same devices and a long prompt arriving mid-stream steals compute from every user currently decoding, which is exactly the stall that makes agent products feel broken.

Separating the pools lets each side be provisioned for its own bottleneck. Tenstorrent's description is explicit that the prefill pool spreads a prompt across many chips using high device-to-device bandwidth, while the decode pool keeps weights and KV cache in DRAM and runs steps independently. The Ethernet links carry the handoff.

Ethernet is the interesting choice. A proprietary scale-up fabric offers more bandwidth and lower latency than commodity Ethernet, and NVIDIA's position in AI compute rests partly on owning that fabric. Tenstorrent is arguing that for this workload — a mixture-of-experts model with 23 billion active parameters and a serving plan that already separates prefill from decode — standard Ethernet is good enough, and the scale-out economics are better because the interconnect is not a second lock-in. That is a plausible argument, not a proven one. It becomes proven when someone publishes tokens per second per dollar across a multi-node deployment, and nobody has.

The arithmetic that makes 428B fit in one rack

The model is 428 billion parameters in total and activates 23 billion of them per token. In BF16 that is roughly 856 GB before any KV cache or runtime overhead. A single Blackhole Galaxy is 32 Blackhole ASICs with 32 GB of GDDR6 each — the 1 TB Tenstorrent quotes — so a BF16 copy would consume about 84% of the system's memory and leave roughly 170 GB for everything else. On a 1-million-token context, with a KV cache that has to survive multiple agent turns, that is not a comfortable margin.

The published download is 252.4 GB. That is the number that makes the deployment fit: it leaves something in the order of 770 GB of the Galaxy's terabyte for KV cache, activations and the speculative-decoding drafter. The mixed-precision scheme is what buys that room — MXFP8 for the shared and attention-side weights, NVFP4 for the routed experts, which are the bulk of the parameters and the part that is only touched sparsely. Sparse attention and dense MLPs confined to the first three layers reduce the memory traffic further.

This arithmetic is ours, on published numbers, not a Tenstorrent claim. But it explains the shape of the announcement: the port is credible precisely because 4-bit routed experts turn an 856 GB model into a 252 GB file that fits on a 32-chip system with room left for long context.

What Tenstorrent did not publish

There is no tokens-per-second figure, no time-to-first-token figure, no concurrency figure and no price for the Tenstorrent-hosted endpoint. For a deployment whose entire premise is that non-NVIDIA silicon can serve a frontier-class open model, those are the four numbers that would settle the argument. Their absence is not evidence that the deployment is slow. It is evidence that the announcement is a capability statement rather than a benchmark result, and readers should file it accordingly.

The only public speed numbers for MiniMax M3.1 Flash describe a different deployment. MiniMax claims a median above 170 output tokens per second and peaks above 200 on its own API. Artificial Analysis, measuring the hosted API rather than Tenstorrent's hardware, records 95.6 output tokens per second with a 1.48-second time to first token. That is roughly a 1.8x gap between the vendor's claim and an independent measurement on a different serving path, which is a useful reminder that "the model is fast" and "this deployment is fast" are separate claims.

What Artificial Analysis does establish about the model is more relevant to the Tenstorrent use case than the headline benchmark. M3.1 Flash scores 43 on the Intelligence Index, costs about $0.22 per completed index task, and scores 16 on AA-Omniscience with a 93% hallucination rate — meaning that when it did not know an answer, it usually produced a wrong one instead of declining. Tenstorrent tuned this deployment on agent traces. Agent loops are exactly where confident wrong answers are most expensive, because the next tool call inherits them. A fast, cheap serving path does not fix that; it makes the verification layer more important.

What to do with this if you are buying inference

For most teams the practical answer this week is "test it, do not migrate to it." The Playground is the cheapest way to see whether the model's behaviour suits your workload, and Tenstorrent's agent-harness integration is the interesting part to probe, because the deployment was tuned for multi-turn tool use rather than single-shot chat. Bring your own traces. The SemiAnalysis AgentX tuning tells you what Tenstorrent optimized for; it does not tell you how your prompts will behave.

Do not put a production path on it until two things exist: a published throughput and latency figure from Tenstorrent, and your own measurement of quality per task. The first is a vendor away from being answered. The second only you can answer.

One honest note on where OrcaRouter sits in this. M3.1 Flash is not in our catalogue; the MiniMax line there stops at MiniMax-M3. So the ways to call the new model today are Tenstorrent Cloud and MiniMax's own API, and we are not one of them. What OrcaRouter does provide is the comparison baseline you need before committing: MiniMax-M3 and more than 200 other models behind a single key, at provider list price with 0% markup, so the question "is the new model actually better for my workload than the one I already run" can be answered without a second contract and a second billing relationship.

The number that would turn a port into a price

Tenstorrent has demonstrated something real: a 428B mixture-of-experts model, quantized to a mixed 4/8-bit format, running both halves of inference on Tensix silicon over Ethernet, with open-source tooling and a published tuning methodology. That is a portability proof, and it is the first one at this model scale that does not route through CUDA.

It has not demonstrated that the port is cheaper, faster, or more reliable than the GPU path. The Galaxy is a 6U, 32-ASIC, roughly 23-PFLOPS BlockFP8 system that listed at $110,000 when it became generally available in April 2026, before you add the supercluster networking this deployment uses. Whether that hardware, running this model, produces a better cost per million tokens than an H100 or an MI300X cluster is the entire commercial question, and the announcement leaves it blank.

The next thing to watch is therefore not another model on the same hardware. It is a tokens-per-second number and a rate card from Tenstorrent, followed by a second silicon vendor doing the same port. When those exist, open weights will have turned into genuine hardware competition. Until then, the accurate description is a credible architectural demonstration with the commercial half still unpublished.