Intel / discrete

Intel Arc Pro B65

32 GB decides what fits. 608 GB/s decides how fast it runs once it does.

Open in the calculator →Best models for 32 GB, on every card that size →

Find one: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Quick answer

The Intel Arc Pro B65 fits 217 of 320 sized models entirely in its 32 GB; the largest widely used one is Kimi-Linear-48B-A3B-Instruct (30.4 GB at Q4_K_M). Memory decides what fits; its 608 GB/s of bandwidth decides how fast it answers.

Largest popular model that fitsKimi-Linear-48B-A3B-Instruct49B params · Q4_K_M · needs 30.4 GB~253 tok/s Faster than you readOpen in the calculator →
Most downloaded that fitsQwen3.8-27B28B params · Q4_K_M · needs 18.6 GB~32 tok/s Faster than you readOpen in the calculator →

At 8,192 tokens of context with the whole model in device memory. Speeds are estimates from memory bandwidth, not benchmarks run on this card.

Intel Arc Pro B65 · 32 GB · 608 GB/s · Q4_K_M where published · 8,192 tokensSpecificationpublished dataDecode speedestimate

217 of 320 fit entirely

217 of the 320 models the engine can size fit entirely in device memory at Q4_K_M where it is published, otherwise the nearest published format, and 8,192 tokens. 16 run with some layers on system memory (32 GB assumed), and 87 do not run at all.

68%of the models the engine can size fit entirely

32GB of device memory

608GB/s memory bandwidth

16run with system memory

87do not run at 8,192 tokens

Featured models · Q4_K_M at 8,192 tokens, including offload

Model	Needs	Verdict	Decode	Calculator
Qwen3.5-2B2.3B parameters	2.32 GB	fits	~384 tok/s	Open →
NVIDIA-Nemotron-3-Nano-4B-BF164.0B parameters	3.46 GB	fits	~178 tok/s	Open →
Qwen3-VL-8B-Instruct8.8B parameters	7.04 GB	fits	~80 tok/s	Open →
Qwen3.5-9B9.7B parameters	7.05 GB	fits	~97 tok/s	Open →
gemma-4-12B-it12B parameters	9.54 GB	fits	~58 tok/s	Open →
gpt-oss-20b21B parameters	13.8 GB	fits	~189 tok/s	Open →
Qwen3.8-27B28B parameters	18.6 GB	fits	~32 tok/s	Open →
Qwen3.6-27B28B parameters	18.6 GB	fits	~32 tok/s	Open →
Kimi-Linear-48B-A3B-Instruct49B parameters	30.4 GB	fits	~253 tok/s	Open →
Qwen3.8-Flash-Next180B parameters	111.6 GB	needs 80.6 GB of system RAM; 32 GB assumed	does not run	Open →
DeepSeek-V4-Flash-Vision-Exp305B parameters	188.1 GB	needs 156.8 GB of system RAM; 32 GB assumed	does not run	Open →
GLM-5.3-Flash321B parameters	198.3 GB	needs 166.6 GB of system RAM; 32 GB assumed	does not run	Open →

These examples are selected from prominent labs using the catalogue’s latest Hugging Face 30-day downloads and repository-creation freshness signal, with newer releases guaranteed a place. Offloaded rows assume 32 GB of system RAM, and a speed is only shown for a row that runs. Fitting in memory is not the same as loading: whether the runtime and version you have supports each architecture and format on this machine has not been tested here. Each “Open” link carries the same model, format, context and RAM into the calculator.

Fits entirely in Intel Arc Pro B65 memory — 217 of 320 sized models

Every model that fitswhole model resident · offload off

The most demanding model that fits is Kimi-Linear-48B-A3B-Instruct at Q4_K_M: 30.4 GB of the 32 GB, leaving 1.63 GB spare.

Fully resident at 8,192 tokens (or the model’s own maximum, where that is shorter), offload off, at 608 GB/s. Each row is a run of the engine for this configuration; the rows start with current, prominent releases and “fits” is memory, not a tested runtime. Older or less prominent models remain available through this search and “Show all”.

Find a model in this list

Showing 40 of 217 models that fit.

Model	Format	Needs	Spare	Decode	Downloads / 30d	Calculator
Qwen3.8-27BQwen · 28B params	Q4_K_M	18.6 GB	13.4 GB	~32 tok/s	6.9M	Open →
gemma-4-26B-A4B-itGoogle · 26B params	Q4_K_M	16.8 GB	15.2 GB	~178 tok/s	13M	Open →
gemma-4-31B-itGoogle · 31B params	Q4_K_M	22.2 GB	9.83 GB	~23 tok/sAbout reading pace	9.9M	Open →
Qwen3.5-9BQwen · 9.7B params	Q4_K_M	7.05 GB	25.0 GB	~97 tok/s	9M	Open →
Qwen3.5-4BQwen · 4.7B params	Q4_K_M	3.98 GB	28.0 GB	~175 tok/s	7.8M	Open →
Qwen3.6-35B-A3BQwen · 36B params	Q4_K_M	23.1 GB	8.90 GB	~256 tok/s	3.3M	Open →
gemma-4-12B-itGoogle · 12B params	Q4_K_M	9.54 GB	22.5 GB	~58 tok/s	1.9M	Open →
Qwen3.6-27BQwen · 28B params	Q4_K_M	18.6 GB	13.4 GB	~32 tok/s	2.5M	Open →
Qwen3.5-2BQwen · 2.3B params	Q4_K_M	2.32 GB	29.7 GB	~384 tok/s	4.9M	Open →
gemma-4-E4B-itGoogle · 8.0B params	Q4_K_M	5.86 GB	26.1 GB	~157 tok/s	4.4M	Open →
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16NVIDIA · 32B params	Q4_K_M	20.3 GB	11.7 GB	~24 tok/sAbout reading pace	530K	Open →
NVIDIA-Nemotron-3-Nano-4B-BF16NVIDIA · 4.0B params	Q4_K_M	3.46 GB	28.5 GB	~178 tok/s	3.4M	Open →
gemma-4-E2B-itGoogle · 5.1B params	Q4_K_M	4.02 GB	28.0 GB	~414 tok/s	3M	Open →
Qwen3.5-0.8BQwen · 873M params	Q4_K_M	1.46 GB	30.5 GB	~841 tok/s	2.6M	Open →
Qwen3-VL-8B-InstructQwen · 8.8B params	Q4_K_M	7.04 GB	25.0 GB	~80 tok/s	14.6M	Open →
Qwen3.5-27BQwen · 28B params	Q4_K_M	18.6 GB	13.4 GB	~32 tok/s	1.9M	Open →
North-Micro-Vision-InstructCohereLabs · 2.5B params	Q4_K_M	2.99 GB	29.0 GB	~253 tok/s	180.7K	Open →
Qwen3.5-35B-A3BQwen · 36B params	Q4_K_M	23.1 GB	8.90 GB	~256 tok/s	1.6M	Open →
granite-4.2-8bIBM · 8.8B params	Q4_K_M	7.49 GB	24.5 GB	~73 tok/s	128.4K	Open →
GLM-4.7-Flashzai-org · 31B params	Q4_K_M	20.4 GB	11.6 GB	~24 tok/sAbout reading pace	1.8M	Open →
granite-4.1-3bIBM · 3.4B params	Q4_K_M	3.72 GB	28.3 GB	~171 tok/s	521.8K	Open →
LFM2.5-2.6BLiquidAI · 2.7B params	Q4_K_M	2.61 GB	29.4 GB	~274 tok/s	108.1K	Open →
Qwen3-VL-4B-InstructQwen · 4.4B params	Q4_K_M	4.72 GB	27.3 GB	~128 tok/s	3.5M	Open →
granite-4.1-30bIBM · 29B params	Q4_K_M	20.4 GB	11.6 GB	~24 tok/sAbout reading pace	301.5K	Open →
Qwen3-VL-2B-InstructQwen · 2.1B params	Q4_K_M	3.05 GB	29.0 GB	~232 tok/s	2.8M	Open →
granite-4.2-3bIBM · 3.7B params	Q4_K_M	3.72 GB	28.3 GB	~171 tok/s	50.6K	Open →
LFM2.5-230MLiquidAI · 230M params	Q4_K_M	1.05 GB	30.9 GB	~1912 tok/s	88.6K	Open →
Qwen3-0.6BQwen · 752M params	Q4_K_M	2.22 GB	29.8 GB	~356 tok/s	29.7M	Open →
gpt-oss-20bOpenAI · 21B params	Q4_K_M	13.8 GB	18.2 GB	~189 tok/s	6.6M	Open →
granite-4.2-30bIBM · 29B params	Q4_K_M	20.7 GB	11.3 GB	~24 tok/sAbout reading pace	35.8K	Open →
granite-4.1-8bIBM · 8.8B params	Q4_K_M	7.49 GB	24.5 GB	~73 tok/s	179.3K	Open →
NVIDIA-Nemotron-3-Nano-30B-A3B-BF16NVIDIA · 32B params	Q4_K_M	20.3 GB	11.7 GB	~24 tok/sAbout reading pace	875.2K	Open →
LFM2.5-VL-3BLiquidAI · 3.1B params	Q4_K_M	2.85 GB	29.1 GB	~274 tok/s	26.7K	Open →
MiMo-V2.6-Distill-Qwen-9BNEWXiaomiMiMo · 9.4B params	Q4_K_M	6.90 GB	25.1 GB	~97 tok/s	14.6K	Open →
Qwen3-4B-Instruct-2507Qwen · 4.0B params	Q4_K_M	4.72 GB	27.3 GB	~128 tok/s	3.7M	Open →
Ling-3.0-tinyinclusionAI · 7.9B params	Q4_K_M	5.85 GB	26.1 GB	~552 tok/s	17.7K	Open →
Qwen3-8BQwen · 8.2B params	Q4_K_M	7.04 GB	25.0 GB	~80 tok/s	10.7M	Open →
Olmo-3-7B-Instructallenai · 7.3B params	Q4_K_M	7.96 GB	24.0 GB	~68 tok/s	481.7K	Open →
Qwen3-4BQwen · 4.0B params	Q4_K_M	4.51 GB	27.5 GB	~128 tok/s	7.8M	Open →
LFM2.5-8B-A1BLiquidAI · 8.5B params	Q4_K_M	6.06 GB	25.9 GB	~427 tok/s	32.4K	Open →

Sized at the format most people actually download, not at FP16. The catalogue ordered by downloads →

Find the Intel Arc Pro B65: Amazon ↗ · eBay (new and used) ↗Store links may pay us a commission. They never decide which card is suggested — the memory arithmetic does.

Under the hood

The specification behind every figure

What the manufacturer publishes for this device, and the pages it was read from.

Manufacturer specification

Memory scope	dedicated
Capacity	32 GB
Published options	32 GB
Bandwidth	608 GB/s
Memory type	GDDR6
Bus width	256-bit
FP32 peak	Not published
Dense matrix peak	Not published without sparsity
Total board power	200 W

Source ledger

Intel Arc Pro B65 Graphics — Product Specifications ↗

Retrieved 2026-09-13

Caveats

Intel publishes peak INT8 throughput for this card rather than a dense half-precision matrix figure, so no compute roof is priced for it and time to first token is withheld rather than estimated.
Local inference on Intel graphics runs through SYCL or Vulkan backends whose overhead this engine has not calibrated separately; the memory arithmetic is unaffected, the speed estimate is the generic one.

Published capacity is a hardware ceiling, not guaranteed free runtime memory. The calculator shows the runtime reserve separately rather than folding it into a single number.

Run the diagnostic on the Intel Arc Pro B65 →