Want to run a 27B large model locally? There's another card in the 32GB VRAM tier, but don't rush to order yet
Want to run a 27B large model locally? There's one more card in the 32GB VRAM tier, but don't rush to place your order yet. Headline version final draft | 2026-10-18 | Body 1188 Chinese characters | Self-check passed (no corner brackets / no em dashes / H2 without numbering / benchmark data in tables / no translationese) Final title: Want to run a 27B large model locally? There's one more card in the 32GB VRAM tier, but first
💡 What You Will Learn
Want to run a 27B large model locally? There's one more card in the 32GB VRAM tier, but don't rush to place your order yet. Headline version final draft | 2026-10-18 | Body 1188 Chinese characters | S
📜 Table of Contents
Want to Run a 27B Large Model Locally? There's a New Card in the 32GB VRAM Tier, But Hold Off on Ordering
Headline version final draft | 2026-10-18 | Body 1188 Chinese characters | Self-check passed (no corner brackets / no em dashes / no numbered H2s / benchmark data in tables / no translationese) Final title: Want to Run a 27B Large Model Locally? There's a New Card in the 32GB VRAM Tier, But Hold Off on Ordering
27B is the threshold. Not enough VRAM and the weights blow up halfway through loading, crashing straight to desktop.
On October 17, the 32GB VRAM tier gained a new card: the ASUS UGen Intel Arc Pro B70 32GB. No launch event, no price, no release date — just a product page quietly posted on the official site.
It has two numbers that look good and one that's awkward. The good ones are 32GB of VRAM and 310W. The awkward one is that cards built on the same chip have been selling in China starting at 8,999 yuan for a while now.
First, How It Differs from the Reference Design
| Item | ASUS UGen B70 32G | Intel Reference B70 |
|---|---|---|
| Core | BMG-G31, 32 Xe2 / 256 XMX | Same |
| VRAM | 32GB GDDR6 (ECC), 256-bit, 608GB/s | Same |
| Compute | 367 TOPS (INT8), 22.94 TFLOPS (FP32) | Same |
| Card power | 310W | 230W (configurable 160-290W) |
| Power input | 1× 12V-2×6 | 1× 8pin |
| Cooling | Dual-slot blower | Dual-slot blower |
| Outputs | 1× HDMI 2.1b + 3× DP 2.1 | 4× DP 2.1 |
| Suggested price | Not announced | $949 (launched March 2026) |
The power row is the one worth talking about. Same BMG-G31 chip: reference is 230 watts, ASUS pushes it to 310 — 80 more watts, a 35% increase. The reference needs just one 8pin; ASUS switches to a single 12V-2×6 connector and piles on phase-change thermal pads, dual ball-bearing fans, and a 304 stainless steel bracket. Those extra 80 watts are there to hold clocks steady under sustained full load. Running models locally isn't a three-minute gaming session — you leave it going for hours. That's a real consideration.
The cost is 310 watts crammed into a dual-slot blower — the noise won't be polite. Four cards stacked together pull 1,240 watts, so you'll need to swap in a 1,600W PSU before anything else.
What 32GB Can Actually Run
This is the key to deciding whether to buy, and the official list is out.
A single card at 32GB has been validated on 17 models, up to the 27B tier: Llama-3.1-8B in FP16, Qwen3-14B and DeepSeek-V2-Lite 16B MoE in FP8, gpt-oss-20b in MXFP4, and the 27B-class Qwen3.6 at Q8 quantization via llama.cpp's SYCL backend.
Benchmarks are posted too, but they're all Intel-platform figures — not unique to this ASUS card — and no batch sizes or quantization details are given:
| Scenario | Result | Source |
|---|---|---|
| Qwen3 8B, 4 concurrent, 32K context | 49.6 token/s | Backend.AI |
| Qwen3 8B, 8K, 16 concurrent (agent workload) | 188.2 token/s | Backend.AI / Lablup |
| GPT-OSS 20B, 32 concurrent | 1,334 token/s | Backend.AI / Lablup |
| vLLM 50 concurrent | 369 token/s | Level1Techs |
| Time to first token (TTFT) | 0.08s | Puget Systems, MLPerf Client |
| Usable KV cache, single card | ~191K tokens | Backend.AI |
| Dual-card Qwen3-Coder 80B | 43.4 token/s | Community testing (GitHub) |
In one sentence, the capability boundary: up to 27B is comfortable, 80B needs two cards, 120B needs four cards and must use MXFP4 quantization.
While we're at it, let's unpack the most misunderstood marketing claim. The official line says "four cards, 128GB, can run 120B" — but four cards don't equal one 128GB card. In its own validation table, the four-card setup runs the MXFP4 version of gpt-oss-120b and the FP8 versions of 72B to 80B. That's "it fits if you squeeze," not "it just runs." Tensor parallelism, quantization method, how the software splits the workload — miss any one and it won't run.
Price Is the Real Crux
ASUS hasn't quoted a price, but the market gives us two reference lines.
The upper line is rising. On August 17, prices jumped simultaneously in three overseas markets: in the US, Newegg's ASRock version went from $999.99 to $1,299.99, up 30%; B&H listed the reference at $1,779; Europe's lowest transaction price jumped from €1,200 to €1,509, up 26%; Korea went from 1.89 million to 2.8 million won, up 48%.
The lower line is domestic stock. On JD.com's self-operated store and through ASRock, Gunnir, and Maxsun channels, quotes land between 9,499 and 11,999 yuan. Gunnir's B70 TF lists at 18,999, typically sells for 11,999, and has hit a historic low of 9,079. The lower-tier B65 also has 32GB VRAM, 608GB/s bandwidth, and 200W power, with an official retail price of 8,999 — it just cuts the Xe2 cores from 32 to 20 and compute from 367 to 197 TOPS.
Converted to price per GB of VRAM, this path runs about 297 to 375 yuan per GB. For the RTX 5090, also 32GB, overseas communities calculate roughly $125 per GB — a fourfold difference in magnitude.
So whether it's worth it depends on what it dares to quote:
Under 12,000 yuan, it's worth it. At the same price you get an extra three-year global warranty and a beefier power and cooling setup — the only things it has over existing stock.
Over 15,000 yuan, it's not worth it. Same chip, same 32GB, same bandwidth — all you're paying extra for is the brand.
If you just want 32GB and your models are 20B or under, the 8,999 B65 is the better deal. Losing 12 cores matters far less for single-user local inference than a 3,000-yuan price gap, because the bottleneck is usually the 608GB/s bandwidth — and both cards have identical bandwidth.
Two More Pitfalls
First, don't expect Ollama to work out of the box. The official FAQ says it itself: standard Ollama has limited Arc support; the easiest path is llama.cpp's Vulkan backend, and the best-performing is the SYCL backend. Yet the first thing most people do after installing the card is ollama run.
Second, this is ASUS's first card to carry the Arc name, but not its first Intel GPU. In 2021 it made the DG1-4G — passive cooling, compatible with only two motherboards, basically a joke. It then skipped both the Alchemist and Battlemage gaming generations. Five years later it's back, and the first card is professional-only — nothing for gamers. It even launched a separate brand: not ROG, not TUF, not ProArt — UGen.
Who Should Buy, Who Shouldn't
Buy it if: you're building a desktop workstation, you specifically need 32GB VRAM, your largest model is 27B, your case fits a dual-slot blower, and you value the three-year warranty.
Don't buy it if: you want to run 70B or above — dual-card starts at 24,000 yuan, and at that budget you should be comparing other options; you only run 14B or under — a 16GB card saves money; your machine lives in the bedroom — a 310W blower isn't built for quiet environments.
Conclusion
There's nothing wrong with the card itself. The problem is it arrived too late.
This chip launched back in March, and Gunnir and Maxsun already had 32GB cards selling at 9,499. ASUS enters now with only channels, warranty, and those extra 80 watts left to play.
As for whether it's worth it — the answer only means something once it dares to quote a price.
If the China model really sells for 12,000, would you pay an extra three thousand for a three-year warranty? Let's talk in the comments.
Alternative Titles (final call by Uncle Peng)
- "Want to Run a 27B Large Model Locally? There's a New Card in the 32GB VRAM Tier, But Hold Off on Ordering" (selected)
- "32GB VRAM, 27B on a Single Card: ASUS's First Intel Pro Card Is Here, But It Won't Dare Name a Price"
- "Reference Is 230W, ASUS Makes It 310W: Same Intel Chip, What Are Those Extra 80 Watts For?"
- "Same 32GB VRAM, Gunnir and Maxsun Start at 9,499 — Is ASUS's Intel Pro Card Worth It?"
- "ASUS Skipped Intel GPUs for Five Years, and Its First Card Back Is AI-Only — Gamers Get Nothing"
Image Suggestions
- Header: UGen Arc Pro B70 render from ASUS's official product page (3/4 view, dual-slot blower front), captioned "Image: ASUS"
- Image 2: ASUS vs. reference spec comparison as an infographic, highlighting the 310W vs. 230W row
- Image 3: Screenshot of the official three-tier validated model list (17 single-card / 11 dual-card / 7 quad-card), captioned "Image: ASUS official product page"
- Image 4: Screenshot of JD.com prices for B70 / B65 (store names blurred), captioned "Image: JD.com, captured 2026-10-18"
- Backup: Overseas price-hike curve (Korea 1.89M → 2.8M won), captioned "Image: Tianji Web"
Source List
- ASUS official product page: UGen Intel Arc Pro B70 32GB GDDR6 (specs / performance table / three-tier validated model list / FAQ)
- Intel official SKU page: Arc Pro B70 Graphics (230W, 8pin, $949, launched 2026-03-25)
- Wccftech: ASUS Finally Builds An Intel Arc Graphics Card… (2026-10-17)
- VideoCardz: ASUS returns to Intel graphics cards after five years with Arc Pro B70 32GB (2026-10-17)
- Hardware Busters: ASUS Arc Pro B70 Is the Company's First Intel Arc Card (2026-10-17)
- Tianji Web: Overseas Prices Soar! Intel Arc Pro B70 Returns to Public Attention Thanks to AI Workflows (2026-08-18, domestic and overseas pricing and overseas price hikes)
- IT Home: Gunnir Arc Pro B65 TF 32G Professional Graphics Card Listed, Official Retail Price 8,999 Yuan (2026-08-17)
- IT Home / Tencent News / Driving China: ASUS Launches Its First Intel Arc Pro B70 Graphics Card (2026-10-17)
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.