What hardware do I need for Strata?

For Qwen3.8-Flash-Next, 64GB of system RAM and a 12GB or larger graphics card. Strata's docs say the model fits when RAM holds the experts plus about 10GB: 34GB of experts for Q2_0 and 50GB for IQ3_S. With 64GB every original size fits. Supported cards are NVIDIA RTX 20 to 50, AMD RX 7000, RX 9000 and AI PRO R9700, Intel Arc on Linux, Apple Silicon M1 or newer, Ryzen AI Max (Strix Halo) and DGX Spark.

How much RAM does Strata need?

For the ISTA-DASLab Qwen3.8-Flash-Next packs: about 44GB for Q2_0, 46GB for IQ2_XS, 53GB for IQ3_XXS and 60GB for IQ3_S (experts plus 10GB). With 48GB, Q2_0 and IQ2_XS fit. Unsloth's 4 to 8 bit files need about 72GB to 136GB. A 32GB PC runs Q2_0, IQ2_XS and the Coder only with a 24GB card, in a low-RAM mode that maps experts from disk.

Does a bigger GPU help Strata?

It makes Strata faster, but Strata's docs say it does not lower the RAM you need, except in low-RAM mode. Community reports put a 12GB RTX 3060 at about 46 tokens per second and a 16GB RX 9070 at about 58 on Qwen3.8-Flash-Next. A 24GB card also runs the picture reader on the GPU instead of the CPU.

Does Strata run on a Mac or a DGX Spark?

Yes. Apple Silicon M1 or newer runs Qwen3.8-Flash-Next; DeepSeek-V4.1-Flash needs 64GB or more and is text only. DGX Spark support arrived in Strata 0.2.7.1 (27 October 2026), Terminal install only, with glibc 2.39 or newer and NVIDIA driver 580.65 or newer. The release notes list 45 to 48 tokens per second on IQ3_S in the first Spark test, and 55 to 67 in the 0.2.9 release check.

Can Strata run DeepSeek V4.1 Flash on a 12GB GPU?

Yes, slowly. Strata's DeepSeek engine needs a 12GB card, 46GB of RAM at least (64GB tested as comfortable), an NVMe SSD and about 265GB of free disk for the 2.00 bpw set. The README lists 9 to 11 tokens per second on an RTX 3060 12GB and 4 to 6 on an RTX 5070 12GB, against 21 to 25 on one Radeon AI PRO R9700 with 251GB of RAM.

← All guides

Guides November 3, 2026

Best Hardware for Strata (2026): Buy 64GB of RAM First

Strata is a one-click local engine for Qwen3.8-Flash-Next and DeepSeek-V4.1-Flash. Its GitHub repo had 38,921 stars on 3 November 2026, and it shipped four releases in the week before. Strata keeps the model's experts in system RAM and puts the rest on the graphics card. So the first thing to buy is RAM, not a bigger GPU.

Bottom Line

Buy 64GB of system RAM first. Strata loads Qwen3.8-Flash-Next’s experts into RAM and the rest onto the graphics card. Strata’s own docs say a bigger card “makes it faster and does not lower the RAM needed.”

Your machine	What Strata runs	Measured speed (community or developer)
32GB RAM + 12-16GB card	Coder pack only (coding)	Slow; experts read from the SSD
32GB RAM + 24GB card	Q2_0, IQ2_XS, Coder (low-RAM mode)	No published number
64GB RAM + 12GB card	Every original size	~46 tok/s (RTX 3060, IQ3_XXS)
64GB RAM + 16GB card	Every original size	48-59 tok/s (RTX 5060 Ti, RX 9070)
Strix Halo 128GB	Every size, plus DeepSeek on Windows	58.9 tok/s (IQ3_S, Strata 0.2.9)

Cheapest upgrade that works: add RAM to the PC you have. A 12GB card is enough to start.
Best new card for it: a 16GB card. It is the tier with the most community reports.
One box, no tower: a Ryzen AI Max+ 395 mini PC with 128GB. Its prompt speed nearly doubled in 0.2.9.
Skip for Strata alone: a DGX Spark. It costs about twice a Strix Halo box and measured in the same speed range.

Ready to buy? See the tested hardware list with current prices.

What Strata Is

	Strata
Repo	Niko1221/Strata, MIT license
Created	24 September 2026
Stars on 3 November 2026	38,921 (4,224 forks)
Releases in the last week	0.2.7 (27 Oct), 0.2.8 (29 Oct), 0.2.9 (31 Oct), 0.2.10 (2 Nov)
Models	Qwen3.8-Flash-Next, Qwen3.6-35B-A3B, DeepSeek-V4.1-Flash-1017
API	OpenAI- and Anthropic-compatible, on localhost
Works with	Claude Code, Codex CLI, OpenCode, Hermes Agent, OpenClaw (`strata connect`)
Install	Desktop app (Windows, Linux, Mac) or `strata launch` in a terminal

Why it trends. It packages one large mixture-of-experts model per family with its own engine, and tunes that engine per card. Hacker News asked “What hardware are you using to run local LLMs?” on 2 November 2026 (47 points), and Strata appears in the replies. Forks for Strix Halo and the RTX 3060 appeared the same week.

RAM Decides What Fits

Strata loads shard 1 of the model at start. Its experts go into RAM. The rest goes onto the graphics card. A second shard, a 29GB lookup table, stays on the SSD.

The rule from Strata’s docs/MODELS.md: RAM must hold the experts plus about 10GB for the OS and other programs.

Pack (ISTA-DASLab)	Experts in RAM	RAM it needs (experts + 10GB)	Strata’s quality note
Q2_0	34 GB	~44 GB	good, fastest
IQ2_XS	35.5 GB	~46 GB	better
IQ3_XXS	43 GB	~53 GB	great
IQ3_S	50 GB	~60 GB	best; matches the full model on the published tests
Coder (IQ1_M, half the experts)	23 GB	~33 GB	coding only

The “RAM it needs” column is our arithmetic from Strata’s rule. So:

64GB: every size fits. IQ3_S leaves little room for other apps.
48GB: Q2_0 and IQ2_XS fit.
32GB: only in low-RAM mode. Setup maps experts from the model files. A 24GB card then runs Q2_0, IQ2_XS and the Coder; a 12-16GB card runs the Coder. Strata warns that most experts then come from the SSD and it is much slower.

Bigger files need much more RAM. Unsloth’s packs of the same model keep the experts at 4 to 8 bits:

Pack (Unsloth)	RAM for all experts	Same top token as Q8_0
UD-IQ4_XS	~72 GB	88.9%
UD-Q4_K_XL	~88 GB	93.1%
UD-Q6_K_XL	~118 GB	94.6%
Q8_0	~136 GB	100% (reference)

For comparison, Strata’s docs put IQ3_S at 87.7% on the same top-token test. NVFP4 packs need about 80GB of RAM and run on NVIDIA cards only. RTX 50 cards have FP4 tensor cores; other NVIDIA cards run NVFP4 at 8-bit precision. See MXFP4 vs NVFP4 for the format.

Plan the disk too. The three smaller Qwen packs download 66 to 76GB, plus about 6GB for the MTP draft layer.

How Fast Each Machine Runs It

Every number below is community-reported or from Strata’s own release notes. None is ours. Versions and settings differ, so read the table as a range, not a ranking.

Qwen3.8-Flash-Next, one request, decode speed:

Hardware	RAM	Pack	Decode	Source
RTX 3060 12GB	64 GB	IQ3_XXS	46.2 tok/s	Strata 0.2.7 notes (default routing)
RTX 4070 12GB	32 GB DDR4	IQ2_XS	48.2 tok/s	Community, modified engine
RTX 5060 Ti 16GB (WSL2)	128 GB	IQ3_S	48.2 tok/s	Community, 0.1.41
RX 9070 16GB	64 GB	IQ3_XXS	58.7 tok/s	Community, 0.1.41
Ryzen AI Max+ 395 (Windows)	128 GB unified	IQ3_S	58.9 tok/s	Strata 0.2.9 notes
Ryzen AI Max+ 395 laptop, 70W	128 GB unified	UD-IQ4_XS	59.2 tok/s (8K)	Community, 0.1.41
DGX Spark	128 GB unified	IQ3_S	45.5-48.5 tok/s	Strata 0.2.7.1 notes
1x Radeon AI PRO R9700 32GB	—	IQ3_S	67 tok/s	Strata 0.2.8 notes
2x Intel Arc Pro B65 32GB	—	IQ3_S	71-74 tok/s	Strata 0.2.8 notes
2x TITAN RTX 24GB	125 GiB	IQ3_S	70.9-77.3 tok/s	Community, 0.1.40.3
Apple M5 Max (memory not stated)	unified	IQ3_S	79-104 tok/s	Strata 0.2.9 release check
2x RTX A6000 48GB	—	IQ3_S	148-169 tok/s	Strata 0.2.9 release check
RTX PRO 5000 72GB	—	IQ3_S	196-212 tok/s (story)	Strata 0.2.7 notes
Tesla V100 32GB	96 GB	UD-IQ4_XS	28.2-32.9 tok/s	Community, source build for sm_70
Tesla P40 24GB x2	125 GB	IQ3_S	35.0 tok/s	Community, source build for sm_61

Read it with four caveats:

The 12GB speed-up is off by default now. Strata 0.2.7 made the RTX 3060 answer 15% faster (46.2 to 53.1 tok/s) by routing to experts already on the card. Version 0.2.8 turned that off again, because users tied it to repeated text in long sessions.
The 8GB numbers in the release notes are not from 8GB PCs. Strata states they come from larger cards limited by hand, on a workstation with a lot of RAM. Only the Intel Arc A750 figures (12.7 to 18.4 tok/s) are from a real 8GB card.
No single RTX 3090 or 4090 number is published for Strata. The two-TITAN-RTX row is the closest 24GB data. We do not estimate one.
The DGX Spark row is the first Spark test. The 0.2.9 release check lists 55 to 67 tok/s on IQ3_S. That is the same range as the Strix Halo box (58.9), not a step above it.

What to Buy

The cheapest route: RAM for the PC you have

If your PC has a 12GB or larger card from the supported list and 32GB of RAM, the upgrade is 64GB of RAM. That moves you from the Coder pack to every original size. A 64GB kit such as the Crucial DDR5 64GB kit fits a DDR5 board. DDR5 prices rose sharply in 2026; check current listings. A DDR4 board needs a DDR4 kit.

The best new card for Strata: 16GB

The 16GB tier has the most community reports: 48 to 59 tok/s with 64GB or more of RAM. The RTX 5060 Ti 16GB is the NVIDIA pick; check current listings for its price. On AMD, the RX 9070 measured 58.7 tok/s.

A 24GB card: for low-RAM PCs and pictures

A 24GB card is the only way to run Q2_0 or IQ2_XS on a 32GB PC. It also moves the picture reader onto the GPU: 0.1 to 0.5 seconds per picture, against a few seconds on the CPU. A used RTX 3090 averaged $1,563 on resale trackers as of October 2026. For text only, 64GB of RAM and a 16GB card is the cheaper path.

One box: Ryzen AI Max+ 395 with 128GB

A Strix Halo mini PC holds every Qwen pack in unified memory, and it runs DeepSeek-V4.1-Flash on Windows. Strata 0.2.9 made its prompts 1.8 times as fast: an 8K prompt went from 791 to 1,424 tok/s. The GMKtec EVO-X2 128GB listed at $3,499.99 as of October 2026. See which Strix Halo mini PC to buy, or start with a Ryzen AI Max+ 395 128GB box. One limit: DeepSeek does not run on Strix Halo under Linux yet.

Not for Strata alone: DGX Spark

The DGX Spark runs Strata since 0.2.7.1. It needs glibc 2.39 or newer, NVIDIA driver 580.65 or newer, and the Terminal install; there is no ARM desktop app. At $6,950 for the 128GB model (November 2026), it costs about twice a Strix Halo box. Strata’s notes put both at 45 to 67 tok/s on IQ3_S for one user. Buy the Spark for CUDA work, not for Strata. If you want a faster single box and already use a Mac, Strata’s 0.2.9 check measured an M5 Max at 79 to 104 tok/s. See best models for DGX Spark.

DeepSeek V4.1 Flash on Strata: Disk and RAM

DeepSeek-V4.1-Flash uses its own Strata engine. What does not fit in memory is read from the SSD, so it needs a lot of disk.

Need	2.00 bpw set
Graphics card	12GB or more (NVIDIA RTX 30-50, R9700 on Linux, Strix Halo on Windows, Mac 64GB+)
RAM	46GB at least; 64GB tested as comfortable
Disk	257GB download, about 265GB free
Context	8,192 tokens, the size it was checked at

Hardware	Output
RTX 5070 12GB, 64GB RAM	4-6 tok/s
RTX 3060 12GB, 46GB RAM	9-11 tok/s
1x R9700, 251GB RAM	21-25 tok/s
4x R9700	74-77 tok/s
Ryzen AI Max+ 395, 128GB	19-21 tok/s code, 11 chat
M5 Max 128GB (Mac, 0.2.8)	18-22 tok/s

The RTX 5070 row is slower than the older RTX 3060 because each token waits on the SSD and the CPU, not the card. A fast NVMe drive matters more than the GPU here. On a Mac, update to 0.2.8 or newer: earlier builds left about 1.3GB of GPU buffers behind per request, and decode fell from 16 to 3 tok/s. For the model itself, see DeepSeek V4.1 Flash GSQ-RCO quants.

Check Your Machine Before You Buy

Strata prints its plan without downloading anything:

strata launch --plan --family qwen --model iq2_xs

The plan shows what goes into RAM, what goes onto the card, and whether setup falls back to the SSD. Run it on your current PC first. If it says the experts do not fit, buy RAM. If it fits and you want more speed, then buy a card.

Sources

Strata repo, README, docs/MODELS.md and docs/COMMUNITY_BENCHMARKS.md: github.com/Niko1221/Strata (read 2026-11-03 via the GitHub API)
Release notes 0.2.5 to 0.2.10, including the 0.2.7.1 DGX Spark section: Strata releases (read 2026-11-03)
Hacker News, “Ask HN: What hardware are you using to run local LLMs?”, 2026-11-02: item 50337431
Prices: price-reference rows for used RTX 3090 (resaleprices.com, 2026-10-30), GMKtec EVO-X2 128GB (vendor store, 2026-10-26) and DGX Spark 128GB (2026-11-02)