AMD Strix Halo/Ryzen AI Max+ 395 for AI – An Honest Review
Credibility score: 63/100 — Mostly Credible. Mixed credibility - some claims are solid, others need verification.
BSmeter analyzed "AMD Strix Halo/Ryzen AI Max+ 395 for AI – An Honest Review" and rated it 63/100 for credibility (a BS score of 37/100 — mostly credible), on 2026-04-27. Its weakest claim — "Device has pros and cons for local AI use" — scored 50/100 and was flagged as opinion. 28 claims were checked against the video transcript. Scores are produced by BSmeter's AI analysis of the transcript, not independent human verification.
Of 28 claims analyzed: 0 scored under 40, 18 between 40 and 69, and 10 at 70 or above.
Claims analyzed
Holds Ryzen Strix Halo / Ryzen AI Max+ with 128GB unified memory — Solid (75/100)
At 0:00
👌
Why this score: Matches known Strix Halo specs.
- Strix Halo is AMD's high-end APU with massive unified memory up to 128GB, as per rumors and leaks confirmed in 2024-2025 announcements.
- GMK Techch Evo X2 and Framework are real devices using these chips; 'Ryzen AI Max+ 395' aligns with Ryzen AI 300/Strix Halo naming.
- No contradictions in web context.
Original quote: “[0:00] In my hand right here is the Ryzen [0:02] Stricks Halo or as you may be familiar [0:05] with the Ryzen AI Max Plus. But [0:08] regardless of what it's called, this is [0:09] the 128 GB of unified system memory [0:13] machine from AMD. Now, the specific [0:15] thing I'm holding in my hand is…”
Framework, GMK Tech, HP have Strix Halo mini PCs — Verified (95/100)
At 0:30
✅👌
Why this score: All named manufacturers confirmed to offer Strix Halo systems.
- GMKtec Evo X2 with Ryzen AI Max+ 395; HP Z2 Mini G1a with Max+ 395; Framework desktop with Max+ 395.
- Matches web context exactly.
Original quote: “of manufacturer that have been putting this inside their systems. Framework has an offering, obviously. GMK Tech has one. There are a bunch of other mini PC manufacturers now putting this in their products. HP even has a little mini desktop with this chip or system inside of it.”
Device has pros and cons for local AI use — Opinion (50/100)
At 0:46
Fair setup — balanced take incoming 👀
Why this score: Subjective assessment of pros/cons in AI performance.
- Speaker promises honest review based on personal testing.
- No specific claims yet; sets expectation for benchmarks later.
Original quote: “for today's video, we're going to be taking a look at the performance of this device with the specific emphasis on using it for local AI. Now, I find truthfully that there are definite pros, but there are some definite cons”
Running Strix Halo on Ubuntu 24.04, skipped Windows — Personal Story (70/100)
At 2:02
Linux swap makes sense for AI nerds ⚙️
Why this score: Personal setup choice aligns with AI use case.
- Framework provides Ubuntu guide for Strix Halo; common for local AI to use Linux.
- Credible anecdote; no contradiction.
Original quote: “So, I have the machine running Ubuntu 24.04. Unfortunately, if you're interested in seeing the performance of this device on Windows, I'm unable to assist being that I actually just immediately put Linux on it. I didn't even boot it into Windows,”
Comparing Strix Halo to dual 3090 Ti rig with 48GB VRAM — Personal Story (60/100)
At 3:00
Fair apples-to-apples test on his rig ⚖️
Why this score: Personal hardware comparison.
- Dual 3090 Ti = 48GB VRAM confirmed viable for 120B models.
- Anecdotal but specific setup; Strix Halo iGPU/NPU vs discrete GPUs is valid contrast.
- No contradiction in web context.
Original quote: “[3:00] OpenAI's open- source GPTO OSS120B, [3:04] show that it worked, and be like, "This [3:06] device is perfect for AI, which is kind [3:08] of partially what we're going to do [3:10] right here. But I want to give an actual [3:12] head-to-head showcase comparison in why [3:15] this is very…”
Dual 3090 Ti gets 13.77 tokens/sec on GPT-OSS-120B — OK (65/100)
At 3:37
Plausible speed for that old iron 🐌
Why this score: Reasonable benchmark.
- 3090 Ti still used for LLMs in 2026; 13.77 t/s fits quantized 120B MoE on 48GB VRAM.
- TPS is standard metric; no direct contradiction but specific setup unverifiable without replication.
Original quote: “[3:37] the exact same OpenAI GPTO OSS120B model on [3:41] the machine behind me and getting a [3:43] speed of 13.77 [3:46] tokens per second in my response.”
Strix Halo runs large LLMs well, using LM Studio — Solid (75/100)
At 3:47
👌
Why this score: Legit demo setup.
- LM Studio is a standard tool for local LLMs per 2026 updates.
- Strix Halo's NPU designed for AI/LLM workloads, matches thesis.
Original quote: “[2:27] flavor of Linux on it. Now, I would like [2:29] to start out by just showcasing one of [2:31] the main strong selling points of the [2:34] system, which is its ability to run [2:36] large LLMs. So, for this, we're going to [2:38] hop into LM Studio. I have heard [2:41] community members…”
Strix Halo hits 46.9 tokens/sec on same 120B model — OK (65/100)
At 4:24
3x faster than dual 3090s? Impressive if real 📈
Why this score: Strong result for APU.
- 46.9 t/s on Strix Halo vs 13.77 on beefy rig suggests excellent NPU/iGPU efficiency for quantized LLM.
- Matches Strix Halo's high-end AI focus; live demo but exact quantization/settings unspecified.
Original quote: “[4:24] which is the token speed. So, this got [4:26] 46.9 tokens per second in running the [4:29] same exact model that the machine behind”
Strix Halo hits 47 tokens/sec vs 13.77 on big desktop machine — Solid (80/100)
At 4:30
👌✅
Why this score: Benchmark aligns with GMKtec EVO-X2 capabilities.
- Web context confirms Ryzen AI Max+ 395 in GMKtec EVO-X2 (April 2025 release) runs large LLMs like 70B/120B models effectively on local hardware.
- Efficiency edge over 1500W multi-GPU setups matches known APU advantages for local inference.
Original quote: “46.9 tokens per second in running the same exact model that the machine behind me, which is significantly larger, more power hungry, um likely louder if it's not liquid cooled and far less efficient. So, 47 tokens per second versus 13.77 tokens per second.”
GMKtec uses 230W brick vs 1500W desktop with 2x450W GPUs — Solid (75/100)
At 5:19
👌
Why this score: Power specs consistent with hardware norms.
- Modern AI mini PCs like GMKtec EVO-X2 have ~90-230W TDPs, vs 1500W+ for multi-GPU rigs (900W+ GPUs alone).
- Matches web context on efficiency focus for Ryzen AI 10000 series NPUs/iGPUs.
Original quote: “the power supply on this GMK tech is around 230 watts of the power brick that actually plugs into the computer. The machine behind me, which just ran the same model at a much slower rate, has a 1500 watt power supply in it with two graphics cards that will each take up to 450 watts of power.”
Best current option for running local LLMs — Opinion (50/100)
At 6:25
Fair take for efficiency chasers — not wrong 👀
Why this score: Subjective assessment of value.
- GMKtec EVO-X2 with Ryzen AI Max+ 395 praised for compact AI/LLM performance in 2025 reviews.
- 'Best' depends on needs (e.g., vs Apple Silicon or Intel alternatives), but efficiency claims hold up.
Original quote: “for running LLMs, this device is probably one of the best current options”
Strix Halo best for running LLMs, Apple also good — Opinion (50/100)
At 6:46
Fair take — Strix Halo's NPU does shine for LLMs 👌
Why this score: Subjective assessment of Strix Halo as top for LLMs. Web context confirms strong NPU and unified memory support up to 128GB, effective for MoE models, rivaling Apple Silicon in some tasks. *Apple remains competitive*, but versatility on Windows/Linux adds value.
Original quote: “for running LLMs, this device is probably one of the best current options out there. Obviously, you have Apple devices with their unified memory setups, and truly that is a perfectly acceptable option as well.”
64GB RAM version at MicroCenter for $1399 — Personal Story (55/100)
At 8:25
$1399 for 64GB? If real, that's an absolute bargain ⚠️
Why this score: Reported pricing observation. Speaker saw 64GB Strix Halo at MicroCenter for $1399 vs their 32GB? at $1799. Web context shows higher prices like Minisforum $2899 for 64GB, but *prices vary by config/manufacturer*; seems low but possible promo/deal.
Original quote: “Now, I will note that at MicroEnter, I noticed they also have the 64 GB RAM version of this device, which they have listed at $1399. And truthfully, if this one is well, $17.99 when I purchased it, but even $18.99, I think it's a no-brainer”
Strix Halo device too expensive vs top-of-line — Opinion (50/100)
At 8:32
Fair take on pricing — depends on your wallet and location 🤷♂️
Why this score: Speaker expresses personal opinion on value, comparing to alternatives like Framework or HP. No specific prices quoted, so subjective. Strix Halo's high-end APU positioning (up to 128GB unified memory) justifies premium for AI users, but accessibility varies by retailer.
Original quote: “this one kind of was just confusing to me because it seemed to be far too expensive comparative to what the topofthe-line device was. And really your options are going to vary depending on whether or not you're close to a microenter, whether or not you want this specific model or you want a…”
AMD support experimental, reviews gloss over real use — Personal Story (65/100)
At 9:07
His weeks of use vs review hype — real talk 👌
Why this score: Anecdotal experience aligns with known AMD AI ecosystem gaps. Web context notes AMD trails NVIDIA in software support (ROCm vs CUDA), requiring community tweaks for non-LLM tasks. Early Strix Halo reviews often focus on LLMs; deeper testing reveals setup friction.
Original quote: “it seems like support for this specific machine is still rather experimental from AMD themselves. ... I found in the reviews that I saw for this specific machine, it very much seem to gloss over actually using it for anything beyond just running GPT OSS in LM Studio.”
Non-LLM AI like Qwen/Stable Diffusion relies on open-source community, feels unfinished — Personal Story (75/100)
At 9:43
Bought it himself, community carrying AMD — spot on 😤✅
Why this score: Personal testimony matches web context: AMD excels in LLMs (GGUF/llama.cpp) but non-LLM generative AI (Qwen-Image, Tongyi Wanxiang/"Juan 2.2", Stable Diffusion) needs *community efforts* like GitHub forks due to experimental ROCm/XDNA support. Strix Halo capable but setup-heavy.
Original quote: “In actually having purchased this myself and used it for the past few weeks, I find that there is a lot here that feels unfinished... Everything that I have seen for running anything that is not a local LLM. So for video generation using Alibaba's Juan 2.2 image generation, stable diffusion,…”
GitHub user made Framework Desktop guide for AI image/video gen — Solid (80/100)
At 10:30
👌
Why this score: Confirmed by Framework's ecosystem.
- Framework Desktop (launched Feb 2025) is marketed for local AI like image/video gen, with active open-source community contributions including GitHub guides and forums.
- Web context notes community sharing projects for AI optimization.
Original quote: “this individual on GitHub. They have an instruction video on how to get this working. They're using the framework desktop, and this is all basically done from someone contributing to the open-source community to actually give owners of this system an option to more simply run video and image…”
Personal experience: AMD Strix Halo AI setup buggy and immature — Personal Story (50/100)
At 11:05
Fair take from hands-on fiddling ⚠️
Why this score: Valid personal testimony. Anecdotal report of bugs in AI setup aligns with community feedback on Framework hardware.
- Web context confirms user reports of 'buggy BIOS/firmware' and slow updates for Ryzen AI systems in 2026.
- No contradiction; reflects real early-adopter pains in new APU ecosystem.
Original quote: “It's tough because I find that could I get this working on my system without these issues that I'm running into right here? Absolutely. Yes. The point is that truthfully I've played with this for a while and it's still bugging out... I won't have to fiddle with an ecosystem that is definitely still…”
Open-source community props up half-baked AMD software support — Opinion (50/100)
At 11:51
Spicy but spot-on community vibe 😤
Why this score: Subjective but grounded assessment. Speaker's frustration with manufacturer support vs. community efforts matches documented patterns.
- Framework provides Linux support but relies on forums/GitHub for AI tweaks (e.g., prompt speed benchmarks).
- 2026 reports echo 'halfbaked' firmware issues, making community 'propping up' a fair critique.
Original quote: “The thing I'm seeing is the open- source community is really propping this system up. We have things like this which are showing performance comparisons in prompt processing speeds with a bunch of different setups on this system... It is just I suppose frustrating to me that a lot of this is still…”
Marketed as all-in-one AI powerhouse but needs workarounds — Opinion (50/100)
At 12:42
Fair gripe — marketing hype vs reality hits hard every time 🚩
Why this score: Speaker expresses personal frustration with software maturity not matching marketing claims. This is subjective opinion, not a factual claim. Web context confirms ongoing issues in AI hardware software integration, like 'AI buyer's remorse' and need for manual setups, aligning with the sentiment without verifying specifics.
Original quote: “My point is I don't know that some of these workarounds or monkey patches should be necessary in a system that has been shown or marketed to us as an all-in-one AI powerhouse solution.”
Reddit/GitHub complaints: no official support vs DGX Spark — Personal Story (60/100)
At 13:02
Echo chamber of gripes checks out for new AI hardware 🙄
Why this score: References specific online complaints as evidence of shared frustration. Anecdotal aggregation from Reddit/GitHub is common for early adopter feedback. Web context supports DGX Spark as a real NVIDIA product with better software stacks, implying competitors like AMD may lag, making this plausible user sentiment.
Original quote: “You can see numerous posts on Reddit. You can see GitHub issues where some folks are saying, "Okay, how does this not have official support?" Someone said, "This is marketed as an Nvidia DGX Spark competitor, but then having zero official software support."”
Roadmap says project unstable, not for production — Solid (75/100)
At 13:58
👌💯
Why this score: Matches industry reality. AI hardware projects often ship with experimental software roadmaps warning against production use. Web context notes similar issues with active development in frameworks like Google's LiteRT and NVIDIA's NemoClaw, confirming instability is standard for emerging AI ecosystems.
Original quote: “the first thing in the road map is basically saying this project is under active development and not stable for production use.”
Did Ubuntu 24.04 unified memory setup but still poor performance — Personal Story (50/100)
At 14:30
Fair troubleshooting tale — Ubuntu memory quirks are real 👌
Why this score: Speaker shares personal experience with Ubuntu 24.04 troubleshooting, including unified memory setup, which aligns with known issues.
- Ubuntu 24.04 has reported memory usage problems in some setups.
- Unified memory is a standard requirement for SoC AI tasks; failure despite steps is plausible anecdote.
Original quote: “There are specific things listed here to try to fix things in Ubuntu 24.04. There is a unified memory setup step that you need to go ahead and do so that the system sees the RAM as like one single available pool. And I've done all of these things and I've put in a decent level of troubleshooting…”
Device fantastic for efficient larger VLMs vs market — Opinion (50/100)
At 15:21
Bold 'fantastic vs everything' take — subjective hype ⚠️
Why this score: Subjective assessment of Strix Halo's efficiency for VLMs/LLMs.
- Matches video thesis on AI focus; on-device AI trend supports efficiency claims generally.
- No specific benchmarks cited here; 'pretty much anything' is exaggerated opinion without direct comparisons.
Original quote: “Like I said, the positives of this device are that it is fantastic at running larger LOMs in a far more efficient manner than pretty much anything on the market.”
AMD engineers actively responding to GitHub issues — Personal Story (65/100)
At 15:45
AMD devs in GitHub comments? That's a green flag 👌
Why this score: Speaker's observation from GitHub of AMD engineers engaging is a credible anecdote for new hardware support.
- Common for first-gen products like Strix Halo to have active vendor responses.
- Builds trust against 'orphaned product' fears; plausible given AMD's AI ecosystem push.
Original quote: “I will say positive things I've noticed is in some of these GitHub issues and places where folks raise issues about support with this system, there does seem to be a decent amount of AMD software engineers who are actually responding to those issues and basically saying like, "Hey, we hear you.…”
Best machine for local LLMs for the price vs Apple — Opinion (50/100)
At 16:31
Fair take — Strix Halo's unified 128GB RAM crushes Apple pricing 👌
Why this score: Speaker's opinion on value proposition aligns with Strix Halo's key strength: unified LPDDR5x-8000 memory up to 128GB, enabling large LLMs (70B models) that Apple's M-series can't match at equivalent spec/price. Subjective but grounded in hardware realities.
- Web context confirms Strix Halo excels for local AI due to memory architecture.
- No direct price benchmarks, but logical given Apple's ecosystem lock-in.
Original quote: “I would say that if you're interested in playing with local AI, you probably can't get a better machine to run local LLMs for the price than you could with this, barring going and getting some form of Apple system. But even then, I believe the equivalent specification would be significantly more…”
This device not locked into specific ecosystem unlike Apple — Solid (80/100)
At 16:46
✅ Open ecosystem win over Apple
Why this score: Accurate comparison. Strix Halo/Ryzen AI Max+ 395 runs on Windows/Linux with standard x86 compatibility, no proprietary lock-in like Apple's ARM/MacOS ecosystem.
- Web context notes flexibility for local AI setups, e.g., Linux with ROCm/Vulkan.
Original quote: “and then you're locked into a specific ecosystem, which this device is not.”
Requires lots of setup/tweaking not advertised, not production-ready — Personal Story (60/100)
At 17:17
Real talk on the Linux tweaks — marketing oversold the plug-and-play 💻⚠️
Why this score: Personal experience valid as anecdote. Speaker bought and used it, notes unadvertised config hurdles (e.g., BIOS/kernel/ROCm for LLMs on Linux), which matches web context on Strix Halo needing specific setups despite power. *Not a universal fact, but honest caveat.*
- Caveat: Their 'mistakes' could be user error, but highlights valid gap between hype and reality.
Original quote: “It is possible that I've made some mistakes in terms of the configurations here or something like that. But the overall point and what I would like to resonate from this video is that the level of setup and tweaking that was necessary that may have led to those mistakes was not something that was…”
See the full analysis with timestamps →