Strata: One RAM Stick Disrupts Three Mini-PC Ecosystems — 10 Days In

Three numbers first. On September 24, a repository called Strata appeared on GitHub. As of today: 9,194 stars. From September 30 to October 3, it shipped 10 versions in four days. The latest, v0.1.34, landed late on October 2 and pulled AMD's entire product line into official support — RX 9070, 9070 XT, 7900 XT/XTX, even the workstation R9700.

What it does in one sentence: it lets a 12GB VRAM gaming card run a 125-billion-parameter model that normally requires a server.

On the official benchmark sheet, an RTX 5070 (12GB) plus 64GB RAM runs Qwen3.8-Flash-Next quantized to Q2_0 at 93 tokens per second — faster than you read. An older 24GB 3090 is estimated at 100–140. Installation is one-click, on both Windows and Linux.

The source material is the transcript of a recent video; I verified every conclusion against the repository. The broad strokes are right, but one technical detail is inverted — more on that below.

Why One RAM Stick Is Enough

Look at the model's shape first. Qwen3.8-Flash-Next is Alibaba's Mixture-of-Experts model open-sourced in late August: 125 billion total parameters, only 6 billion actually used per token. The remaining 95% of weights sit idle most of the time.

Idle weights have one property: they don't need to be fast, they just need to fit.

Strata is built entirely around that sentence. The most frequently routed experts go in VRAM, less-used ones in RAM, and cold ones stay on the SSD — the router predicts which experts the next token needs and prefetches that layer from disk. The result: 12GB of VRAM suffices, because only the activated subset is ever computing; the 100+ GB of total weights are spread across RAM and SSD, both far cheaper.

The video's error was the phrase "cold KV cache goes into memory." The second half is right — some weights are indeed placed in RAM; the first half is wrong — the KV cache essentially stays in VRAM and can be quantized to save space. What actually gets tiered around is the expert weights. This isn't nitpicking: the SSD is the fourth storage tier of this scheme, and the reason an 111GB Q4-quantized model can run on a 64GB-RAM machine is streaming cold experts straight from disk plus prefetching. Underestimate the SSD tier and you underestimate how big a model this can hold.

Three Boxes, Three Situations

The video went through AMD, NVIDIA, and Apple mini-PCs one by one. Their situations do differ, but the underlying logic is more interesting than intuition.

AMD is hurting most — hurting mathematically. The AI Max+ 395 mini-PC's unified memory caps at 128GB, while quantized flagship open-source MoEs — DeepSeek V4 Flash 284B, GLM-5.3-Flash 320B — simply don't fit. That's not an optimization problem, it's a capacity problem. It runs Qwen3.8-Flash-Next fine (the community has a dedicated halogen-flash-server, 831 stars), but the flagship door, 128GB can't get through. One commenter put it more bluntly: he bought an AI Max+ 395 for ~15,000 yuan, and at current prices he wouldn't buy again.

NVIDIA's DGX Spark isn't in the blast radius. Starting at ~33,000 yuan, with dual units daisy-chained over 200G networking to run bigger models, and a complete ecosystem. But what it's losing isn't existing users — it's the upgrade path: previously, if your VRAM fell short, Spark or a mini-PC was the only option; now a single RAM stick bypasses that. Between 33,000 yuan and a 1,000-yuan RAM stick, it's the middle band of demand that's been intercepted.

Apple is best positioned, but "least affected" doesn't mean "unaffected." The M series' unified-memory bandwidth advantage remains — Strata's tiering is genuinely not equivalent to unified memory, as the video honestly admits. But the capacity wall hits both: the M series caps out too, and can't fit 320B either. Someone on GitHub has already ported the same SSD-paged-expert approach to Apple Silicon (amperor-strata) — Apple hasn't escaped the blast; the shockwave just hasn't arrived yet.

| Route | Representative config | Qwen3.8-Flash-Next | Price | |---|---|---|---| | Strata tiering | 12–24GB GPU + 64GB RAM | 93 t/s (measured on 5070) | GPU + RAM, a few thousand yuan | | AMD unified-memory mini-PC | AI Max+ 395, 128GB cap | Flash-Next runs; 284B/320B won't fit | 15,000–20,000+ yuan | | DGX Spark | 128GB unified memory, dual-unit linkable | Flagships run; full ecosystem | 33,000 yuan per unit |

The Cards That Won

The boxes took the hit; the winners are the GPUs sitting in your drawer.

The biggest winners are people holding 16GB unmodified NVIDIA cards: 5070 Ti, 5080, 4080. The more VRAM, the more resident experts, and the closer to "it actually fits." This demographic was previously the mini-PC's prospective buyer base — the video says they can now skip the purchase, and I agree.

AMD cards are playable, but "running great" needs a couple more days — 9070/7900 XTX only entered the official list on October 2, and efficiency is still climbing. One number worth watching: the RX 6800 went from 31 to 42 tokens/sec after v0.1.37 fixed its VRAM budgeting — a 35% jump in one version. The direction is real, and so is the gap: AMD's driver and memory-offload efficiency penalty versus NVIDIA is larger, and the video didn't exaggerate that.

The most surprising revival is the 3060 12GB. RTX 30-series 12GB sits right at the official support floor. "The 3060 doing serious work" used to be a joke; now a machine with 32GB of RAM can run the lowest quantization tier. There's room below that too: v0.1.38 on October 3 let 6GB VRAM cards boot, and RTX 20-series code quietly landed — a community member showed an RTX 2070 8GB plus 64GB RAM hitting 65 tokens/sec (single source; no second confirmation yet).

Which Gear the Flywheel Turns On

The video ends by saying online models are feeding back into local deployment — "the flywheel effect is visible." The judgment is right, but worth taking the gears apart.

It's not some mysticism of "models got smarter so the community got busier." The real gear is architecture: GLM-5.3-Flash is 320B with 18B active, DeepSeek V4.1 Flash is 552B with only 8B active on input, Qwen3.8-Flash-Next is 125B with 6B active. All three flagships are driving the activation ratio down. The smaller the active set, the less the question "how big a card runs how big a model" depends on VRAM — and the more it depends on whether you're willing to buy RAM and SSD. The sparse architectures hammered out in the online-model arms race, released as open weights, are directly becoming fuel for local deployment.

Downstream feedback speed tells the same story. NInfer, a single-GPU inference engine with 2,673 stars, was ported by the community to V100, 4090, and Windows within a week — the V100 port runs quantized Qwen3.8-27B at 228 tokens/sec. Using online large models to write code, patch frameworks, and do ports, with the output landing on local hardware. Every link in that loop is a publicly verifiable repository.

The video's line that "the 3060 12GB can now openly do serious work" — checked, and it holds.

One final detail the video missed. v0.1.38 also fixed a security issue: local services without an API key set no longer respond to cross-site requests from other websites. As local inference is increasingly run as a serious service, security is starting to be taken seriously too — this repository is maturing at the same pace it ships.

---

*Verification notes: Strata repo README and all release notes v0.1.29–38 (GitHub, 9,194★ as of 10-04); Qwen3.8-Flash-Next 125B-A6B+51B Engram (HuggingFace/GitHub/lmsys, multiple sources); GLM-5.3-Flash 320B-A18B (z.ai official/docs, multiple sources); DeepSeek V4 Flash 284B-A13B, V4.1 Flash 552B with 8B active input / 16B output (multiple sources); DGX Spark $3,999 USD, from 32,999 yuan in China; AI Max+ 395 systems 15,000–21,999 yuan (forum comments / HP Z2 Mini pricing); RX 6800 31→42 t/s and AMD support list (release notes text); RTX 2070 8G 65 t/s is a single-source community report, flagged as such.*