AMD Medusa and Local AI: What's Confirmed for 2027
AMD's own blog names "Medusa" as its 2027 AI PC family and targets over 10x AI compute growth since the first Ryzen AI chip. Here's what that does and doesn't tell you about running LLMs locally.
Key takeaways
- AMD confirmed the name and target on its own blog: "Medusa" is the 2027 successor in its AI PC lineup, with a stated goal of over 10x AI compute growth since the first-generation Ryzen AI chip.
- That figure is a cumulative AI-compute target (CPU+GPU+NPU), not a tokens-per-second multiplier for LLM inference.
- Medusa is not the Ryzen AI Max+ 395 (Strix Halo) you can buy today — it's the next platform in the same family, with no published memory spec yet.
- Tech press, citing AMD's own technical documentation spotted online, reports a mobile "Medusa Point" variant scaling up to 22 cores on the FP10 platform — reported, not a public spec sheet.
- What decides LLM usefulness is memory capacity and bandwidth, not NPU TOPS, and neither is confirmed for Medusa yet.
What AMD actually confirmed
Your private ChatGPT, free, on your own machine in an hour — LM Studio, Ollama, Open WebUI, your documents, no cloud.
- Lifetime online access
- PDF + files
- 30-day refund
In a blog post published by AMD itself, the company places "Medusa" as the next step in its client AI platform lineage — Phoenix, Hawk, Strix, Gorgon, then Medusa in 2027. AMD states it expects "an expected >10x performance increase from the first-generation Ryzen AI to the upcoming 'Medusa' family in 2027," and describes each generation as pushing "performance per watt, NPU capability, and efficiency."
That blog confirms a name, a year, and a compute-growth target. It does not detail CPU architecture, GPU generation, memory configuration, or pricing for Medusa.
Medusa isn't the Ryzen AI Max+ 395
A common mix-up: the Ryzen AI Max+ 395 (codename Strix Halo), already shipping and documented on this site with 128GB of unified memory, belongs to the current generation. Medusa succeeds it, but nothing in AMD's own source guarantees an equal or larger memory ceiling — assuming that would be extrapolation, not fact, until AMD publishes a memory spec for Medusa.
Tech press (WCCFTech), citing AMD's own internal technical documentation spotted online, reports a mobile "Medusa Point" variant scaling up to 22 cores on the FP10 platform. That's closer to a leak of internal documentation than a public AMD spec sheet with a full product page. Other Medusa-branded variants for desktop and a higher-end "Halo" tier circulate in the same press coverage, but their exact configuration is, again, unconfirmed by any AMD-signed primary source.
Timing context: Intel's competing "Nova Lake" client platform is also being discussed for a late-2026-into-2027 window. The two roadmaps overlap loosely, but neither company has locked a firm consumer ship date, so treat them as parallel bets rather than directly comparable launches.
AMD Medusa for local AI: state of play
| Point | Status | Source |
|---|---|---|
| Name "Medusa" and 2027 timeframe | Confirmed by AMD | AMD's own blog |
| >10x AI compute target vs. first Ryzen AI | Confirmed by AMD (stated goal) | AMD's own blog |
| Mobile variant up to 22 cores (FP10) | Reported via internal documentation | WCCFTech |
| RDNA 5, detailed Zen 6, AM5 socket | Not confirmed by an up-to-date primary source | To verify at launch |
| RAM capacity, LPDDR6, bandwidth | Unconfirmed | Do not anticipate |
| ROCm / llama.cpp support for Medusa specifically | Untestable before release | First independent benchmarks |
What actually matters for a local LLM
A cumulative AI-compute growth figure (">10x") describes projected compute capacity, not a tokens-per-second multiplier for an LLM. Local inference is still dominated by memory bandwidth: without published memory capacity and bandwidth for Medusa, no throughput estimate today is reliable. See our unified memory explainer for why this matters more than raw compute figures.
The Ryzen AI Max+ 395's usefulness for local AI comes less from its NPU than from unified memory shared between CPU and GPU, which lets it load large models. If Medusa keeps that architectural choice with more memory or higher bandwidth, it will genuinely matter for local AI; if not, the advertised AI-compute growth will mostly help lightweight NPU tasks like video effects, not large-model inference. Our why VRAM matters more than TFLOPS guide covers the same logic on the GPU side.
10x AI compute is not 10x tokens per second. AMD's target covers cumulative AI compute across CPU, GPU and NPU over the whole product trajectory since the first Ryzen AI chip — not a measurement of generation speed with a real model. Treat it as a marketing target until independent benchmarks exist.
ROCm and what software support implies
The current Ryzen AI Max+ 395 is already supported by ROCm and by llama.cpp builds with AMD GPU acceleration, following the same setup path documented in our Ollama on ROCm install guide. That's a positive signal for Medusa within the same product family, but no source confirms certified ROCm support for Medusa ahead of its actual release — every new AMD GPU architecture has historically needed months of ROCm updates after hardware launch before support matured. The NPU side follows a similar pattern, as our NPU explainer covers: a capable NPU on paper still needs a software stack that recognizes it before it does anything for local inference.
Buy now or wait for Medusa?
- Need a local LLM machine now: the Ryzen AI Max+ 395 ships today with a documented 128GB of unified memory. That's a measurable choice, not a 2027 promise.
- Project timeline 2027–2028: wait for full Medusa spec sheets before deciding, especially on the memory capacity OEMs actually ship.
- Care mainly about NPU-driven local AI: NPU compute barely helps large-model inference — memory and its bandwidth decide that, not advertised TOPS.
- Deciding based on the "10x" headline: that's a cumulative AI-compute target AMD announced, not a measured text-generation speed.
Sources: AMD's official blog on its AI compute trajectory, WCCFTech on the Medusa family.
Frequently asked questions
Has AMD Medusa launched?
No. AMD confirmed the name and the 2027 timeframe on its own blog, along with an AI-compute growth target, but has not published a full spec sheet or an exact ship date.
Is Medusa the successor to the Ryzen AI Max+ 395?
Yes, in AMD's platform lineage, but no official memory spec guarantees a capacity equal to or greater than the 128GB on today's Ryzen AI Max+ 395.
Will Medusa use LPDDR6?
Not confirmed by an AMD primary source as of this writing. Don't assume this memory choice before an official announcement specific to Medusa.
Does AMD's "10x" claim mean 10x faster for LLMs?
No. It's a cumulative AI-compute target (CPU, GPU, NPU) across the product trajectory since the first Ryzen AI chip, not a measured tokens-per-second figure for a given model.
Should I wait for Medusa to build a local LLM machine?
Only if your project can wait until 2027–2028. For an immediate need, today's Ryzen AI Max+ 395 remains the measurable choice, already supported by ROCm.
A current option for local AI: GMKtec EVO-X2 64GB / 1TB (Ryzen AI Max+ 395). Match memory to your model and software. A mini PC is a complete PC alternative; Mac/MLX and CUDA instructions require compatible hardware.
Amazon Check GMKtec EVO-X2 64GB / 1TB (Ryzen AI Max+ 395) price →As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.
Found an error or have feedback? Let us know — it helps everyone who reads this guide.