--- title: Day-zero model support is the new chip-war front: MetaX and Tenstorrent ship MiniMax M3.1, AMD publishes Instinct guidance description: MetaX and Tenstorrent both claimed day-zero support for MiniMax's new M3.1 Flash while AMD published Instinct guidance for M3.1. The accelerator war's newest battleground is how fast a new model runs on your chip. canonical: https://hangingcontext.com/stories/2026-10-24/day-zero-chip-support/ last-updated: 2026-10-24 url: https://hangingcontext.com/stories/2026-10-24/day-zero-chip-support/ published: 2026-10-24T01:06:29Z edition: 2026-10-24T00 newsroom: artificial-intelligence beat: infrastructure trust: confirmed source_count: 4 cited_manifest_ids: ['1792724469043796886', '1792729736823688086', '1792750803362306749', '1792729915511228506'] --- # Day-zero model support is the new chip-war front: MetaX and Tenstorrent ship MiniMax M3.1, AMD publishes Instinct guidance *MetaX and Tenstorrent both claimed day-zero support for MiniMax's new M3.1 Flash while AMD published Instinct guidance for M3.1. The accelerator war's newest battleground is how fast a new model runs on your chip.* Confirmed · 4 sources · Artificial Intelligence · Published 2026-10-24 01:06 UTC ## Key facts - MetaX says its XiYun C-Series GPUs completed Day 0 adaptation for MiniMax M3.1 Flash via its self-developed MXMACA stack — its 45th such day-zero port since December 2025 — and that MXMACA is natively compatible with vLLM and SGLang. (self-reported) [cites: 1792724469043796886] - Tenstorrent says it serves MiniMax M3.1 Flash on a supercluster of Blackhole Galaxy systems with prefill and decode split, tuned on SemiAnalysis AgentX traces, at $0.15 per 1M input tokens and $0.60 per 1M output — 70% lower agent-task cost than M3. (self-reported) [cites: 1792729736823688086] - AMD published deployment guidance for MiniMax-M3.1 — the non-Flash model, a distinct release — on Instinct MI350 GPUs, with an optimization stack of ROCm, vLLM, ATOM, and optimized AI kernels. (self-reported) [cites: 1792750803362306749] - SiFive argues that scaling AI clusters to 1.6T makes physical interconnects the datacenter's pressing constraint, and positions configurable RISC-V IP — three families for real-time control, accelerator control, and management — for NICs, switches, and optical modules. (self-reported) [cites: 1792729915511228506] ## A two-horse race to day zero Two chip companies are racing to run MiniMax's new M3.1 Flash on day zero. MetaX says its XiYun C-Series GPUs completed Day 0 adaptation for the model — inference running on its hardware the same day the weights were released — through its self-developed MXMACA software stack. Tenstorrent says it is now serving M3.1 Flash on its cloud on a supercluster of Blackhole Galaxy systems, with prefill and decode split across machines. The subtext: in the accelerator wars, the scoreboard is no longer just throughput or price — it is who runs the newest model first. [cites: 1792724469043796886, 1792729736823688086] MetaX's own numbers tell you how routine the company wants this to look: it says it has completed day-zero adaptation for 45 flagship models since December 2025, spanning MiniMax, Zhipu AI, StepFun, DeepSeek, and Alibaba Qwen, and that MXMACA is natively compatible with the mainstream inference engines vLLM and SGLang. Tenstorrent's pitch is different — not just running the model, but serving it with a tuned stack. The company says it tunes its deployment on SemiAnalysis's AgentX coding-agent traces and splits prefill and decode across its Galaxy infrastructure, targeting production agent workloads rather than launch-day benchmarks. [cites: 1792724469043796886, 1792729736823688086] The model they are both chasing is built for agents. MetaX describes M3.1 Flash as designed for agent scenarios, Computer Use, and software engineering, with a 1M-token context window, image and video input understanding, and adjustable reasoning depth. Both companies put the economics front and center: MetaX says comprehensive token usage cost for agent tasks dropped 70% versus the previous generation, and Tenstorrent lists $0.15 per million input tokens, $0.60 per million output, and $0.015 per million cached input. Read those as vendor claims, but note the pattern — the day-zero race is being fought over agent workloads, not chatbot demos. [cites: 1792724469043796886, 1792729736823688086] > **From Synorb** — Claims with a source link. A Manifest holds a Signal of structured claims and a Brief, plus the source it came from. Your agent can check a sentence against it. [Read the guide](https://synorb.com/connect/start) ## Flash versus M3.1: keep them straight Now the distinction that matters, because five characters are doing a lot of work: AMD published guidance this week too — but for MiniMax-M3.1, not M3.1 Flash. Its post walks developers through running M3.1 on Instinct MI350 GPUs with ROCm, vLLM, ATOM, and optimized AI kernels. Two different model names, landing within days of each other, with different deployment stories. Anyone evaluating stacks should treat M3.1 and M3.1 Flash as distinct models until a vendor explicitly says otherwise; the names are close enough to blur by accident. [cites: 1792750803362306749] The Flash variant's internals, per the vendors: Tenstorrent says MiniMax replaced global grouped-query attention with sparse attention, and that MLP weights, attention weights, indexer, and shared expert were quantized to MXFP8 while routed expert weights went to NVFP4 — the kind of aggressive quantization that explains the cost numbers. On scores, the companies cite 76.2 on BankerToolBench (Tenstorrent says that ties Claude Opus 5), 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82% on CyberGym. Vendor-reported benchmarks are marketing with measurements attached; the architectural details are the part an executive can take to their own hardware team. [cites: 1792729736823688086, 1792724469043796886] ## Below the race, the wiring Underneath the model-port race, the other front is the wire. SiFive's post argues that as AI clusters scale toward 1.6T interconnects, physical interconnects — not GPUs — are becoming the datacenter's most pressing constraint. The company's case: copper traces face severe reach and signal-integrity penalties at 1.6T speeds, 800G optical modules must fit roughly 8 to 18W power budgets, and standards like UALink 2.0 and the DC-SCM specification are moving management and security onto standard modules. Its answer is configurable RISC-V cores — three IP families aimed at real-time control, accelerator control, and management — positioned for NICs, switches, and optical modules. The chip war is not just whose silicon runs the model; it is whose wiring keeps the cluster fed. [cites: 1792729915511228506] Step back and the week's chip news reads as one trend: the competitive surface has moved up and down the stack at once. Up, toward same-day model support as a differentiator that developers can feel; down, toward the interconnect plumbing that decides whether a cluster of accelerators performs like one machine. For a tech executive, the implication is that accelerator vendor selection is becoming a software-and-support question — stack maturity, day-zero track record, and whose engineers show up when the weights drop — at least as much as a silicon question. That is a procurement criteria change, not just a press-release cycle. [cites: 1792724469043796886, 1792729736823688086, 1792729915511228506] ## What is still unresolved - Whether MetaX's first-to-day-zero claim and Tenstorrent's serving claims can be independently verified. - Whether AMD will publish equivalent guidance for the Flash variant of MiniMax-M3.1. - Whether day-zero porting speed converts into durable developer lock-in, or remains a press-release contest. ## Decision card - **Useful for:** Tech executives tracking alternative-accelerator readiness and the software-maturity side of chip vendor selection - **Not for:** Procurement decisions on the numbers alone — every figure here is a vendor claim, not an independent benchmark - **Compare against:** Your own vLLM or SGLang deployment costs and latency on your current fleet ## Provenance ledger | Manifest id | Title | Source | Claims used | Published | |---|---|---|---|---| | `1792724469043796886` | 沐曦股份曦云C系列GPU Day 0适配MiniMax M3.1 Flash | [metax-news](https://metax-tech.com/ndetail/12678.html) | 1792725965711169440, 1792725965775748268, 1792725965771671797, 1792725965735148997, 1792725965783565172, 1792725965727439967, 1792725965743120150, 1792725965751966197, 1792725965759403656, 1792725965767051440, 1792725965791252341, 1792725965799892135, 1792725965807109185, 1792725965815680890, 1792725965823290417 | 10-23 00:00 | | `1792729736823688086` | MiniMax M3.1 Flash, Now Live on Tenstorrent | [tenstorrent-newsroom](https://tenstorrent.com/newsroom/minimax-m31-flash-now-live-on-tenstorrent) | 1792730065638609419, 1792730065641245673, 1792730065601895380, 1792730065603116028, 1792730065606475976, 1792730065599525529, 1792730065608598979, 1792730065627812642, 1792730065632922510, 1792730065635253697, 1792730065623196805 | 10-22 00:00 | | `1792750803362306749` | MiniMax-M3.1 on AMD Instinct GPUs | [amd-blog](https://amd.com/en/developer/resources/technical-articles/2026/minimax-m3-1-on-amd-instinct-gpus.html) | 1792751173144117564, 1792751173147731316, 1792751173149975776, 1792751173151534925, 1792751173152386252 | 10-22 00:00 | | `1792729915511228506` | Beyond the CPU and GPU: SiFive in Datacenter Infrastructure | [sifive-blog](https://sifive.com/blog/beyond-cpu-gpu-sifive-in-datacenter-infrastructure) | 1792730180760515655, 1792730180780879493, 1792730180767296608, 1792730180763734039, 1792730180772317348, 1792730180775863776 | 10-22 00:00 | ## Run these Manifests yourself Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them. 1. **Get a Synorb Key.** Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{"client":"your-agent-name"}`. The key is the `api_key` field of the answer. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; 3 of them are starter Beacons. 2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer `. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief. 3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them. Cited Manifest ids: - `1792724469043796886` - `1792729736823688086` - `1792750803362306749` - `1792729915511228506` ```sh for id in 1792724469043796886 1792729736823688086 1792750803362306749 1792729915511228506; do curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id" done ``` > **From Synorb** — Ask what Synorb covers. The MCP tool synorb-stream-search answers questions like "do you have this publisher?" without using any Manifest allowance. [Connect and ask](https://synorb.com/connect/start) ## Related stories - [Sovereign AI's New Frontier Market Is Kazakhstan](https://hangingcontext.com/stories/2026-10-24/sovereign-ai-kazakhstan-frontier/) - [The chip supply chain is getting auditable: Intel provenance, open chiplets, day-one model ports](https://hangingcontext.com/stories/2026-10-23/silicon-supply-chain-auditable/) - [The AI bill is coming due, and teams are building a control plane to pay it](https://hangingcontext.com/stories/2026-10-23/ai-bill-control-plane/) - [The agent trust stack: retrieval bottlenecks, authorization gaps, and memory that never forgets an attack](https://hangingcontext.com/stories/2026-10-22/agent-trust-stack/) ## Links - HTML: https://hangingcontext.com/stories/2026-10-24/day-zero-chip-support/ - JSON: https://hangingcontext.com/stories/2026-10-24/day-zero-chip-support/index.json - Markdown: https://hangingcontext.com/stories/2026-10-24/day-zero-chip-support/index.md Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify. --- Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-10-24/day-zero-chip-support/. Terms: https://hangingcontext.com/terms/reuse/