In this article
Short answer
TeamGroup T-Create Expert 128GB (2x64GB) DDR5-6000 CL40 is the best kit for running large local LLMs from system memory. TeamGroup lists 128 GB in two modules at DDR5-6000 with XMP 3.0 and EXPO profiles at 1.25 V. Crucial's Pro kit offers the same capacity at a plain 1.1 V, and G.Skill's 96 GB Ripjaws suits Intel buyers who need less.
- 128 GB is the size where a 70B model at Q8_0, a 75 GB file in the Ollama library, loads with space to spare.
- 128 GB still does not hold the biggest open models: Qwen3 235B at Q4_K_M is 142 GB.
- Choose two large modules over four small ones, because a fully populated board runs at a lower supported speed on current AMD chips.
- CPU generation speed follows memory bandwidth, so a model that spills into RAM runs far slower than one in VRAM.
- Best overallTeamGroup T-Create Expert 128GB (2x64GB) DDR5-6000 CL40
- Best 128 GB kit at 1.1 VCrucial Pro 128GB (2x64GB) DDR5-5600 CL46
- Best 96 GB kit for IntelG.Skill Ripjaws S5 96GB (2x48GB) DDR5-6400 CL32
Comparison at a glance#
| Model | Capacity | Speed | CAS latency | Profile | Voltage |
|---|---|---|---|---|---|
| TeamGroup T-Create Expert 128GB (2x64GB) DDR5-6000 CL40 | 128 GB (2 x 64 GB) | 6000 | CL40 | XMP 3.0 / EXPO | 1.25 V |
| Crucial Pro 128GB (2x64GB) DDR5-5600 CL46 | 128 GB kit (64 GB x 2) | 5600 | 46 | XMP 3.0 and EXPO | 1.1 V |
| G.Skill Ripjaws S5 96GB (2x48GB) DDR5-6400 CL32 | 96 GB (48 GB x 2) | Up to 6400 (tested XMP) | 32-39-39-102 | Intel XMP 3.0 | 1.35 V (XMP) |
How we picked#
The guide is limited to kits of 96 GB and above in two modules, because those are the sizes that make CPU offload of large models practical. We compared maker datasheets and product pages for capacity, speed, timings, profiles, module ranks and warranty. Platform limits come from AMD and Intel specification pages. Speed and channel effects come from Maxim Saplin, InsiderLLM, Puget Systems and Phoronix. Model sizes come from the Ollama library. Corsair's article on DRAM for AI informed the capacity thresholds.
The picks#
TeamGroup T-Create Expert 128GB (2x64GB) DDR5-6000 CL40
TeamGroup- Capacity128 GB
- TypeDDR5
- Speed6000
- CAS40
TeamGroup rates this kit at DDR5-6000 with CL40 timings at 1.25 V, which makes it the fastest of the two 128 GB kits here. It ships with one profile type for each platform. Platform support covers Intel 800 and 700 series and AMD 800 and 600 series, and the maker gives a bandwidth figure of 48,000 MB/s.
DDR5-6000 sits above the DDR5-5600 that AMD lists for two dual-rank DIMMs on the Ryzen 9 9950X, so on that platform the profile is an overclock. Intel gives DDR5-6400 for the Core Ultra 9 285K, which leaves the kit within the platform's listed speed. TeamGroup classes both profiles as overclocking, says the result varies with the board, processor and BIOS, and asks buyers to check the board vendor's QVL, its list of tested memory, before buying. It also warns against mixing kits.
On-die ECC is listed for the series, and TeamGroup lists a lifetime warranty. Without a profile the kit runs at the SPD default, such as DDR5-4800 or lower.
| Spec | Detail |
|---|---|
| Capacity (kit) | 128 GB (2 x 64 GB) |
| Type | DDR5 |
| Speed (MT/s) | 6000 |
| CAS latency | 40 |
| Voltage | 1.25 V |
| Profile | XMP 3.0 / EXPO |
| Compatibility | Intel 800/700 series; AMD 800/600 series |
| Warranty | Lifetime |
Pros
- DDR5-6000 rating, the fastest of the two 128 GB kits
- One kit covers both the Intel and AMD profile types
- 1.25 V, lower than the 1.35 V of the G.Skill kit
- Lifetime warranty listed by TeamGroup
Cons
- Needs 1.25 V, above the standard 1.1 V
- On AMD, DDR5-6000 is above the listed DDR5-5600 for two dual-rank DIMMs
Best for buyers who want one kit for large-model offload on either platform
Not for anyone who needs more than 128 GB for the very largest models
Sources: Maker: teamgroupinc.com (opens in a new tab), Benchmark: dev.to (opens in a new tab)
Crucial Pro 128GB (2x64GB) DDR5-5600 CL46
Crucial- Capacity128 GB kit
- TypeDDR5
- Speed5600
- CAS46
Crucial Pro runs at the standard DDR5 voltage of 1.1 V. Crucial lists DDR5-5600 at 46-45-45 as its main profile and DDR5-5200 at 42-42-42 as a second. Both XMP 3.0 and EXPO are supported. That makes it a calm, low-voltage choice.
The trade is timing. CL46 is looser than TeamGroup's CL40. In practice the gap is small for AI, because token generation on a CPU waits on bandwidth rather than latency. Gaessler's llama.cpp notes make that point, and Saplin's tests show speed moving with the memory rate.
Crucial lists a limited lifetime warranty, with 10 years in Germany and France. Crucial also sells a 128 GB kit without the Pro name, part number CT2K64G56C46U5, whose page lists no XMP or EXPO profile, so it runs on JEDEC settings. Check that the part number is CP2K64G56C46U5 before you buy.
| Spec | Detail |
|---|---|
| Capacity (kit) | 128 GB kit (64 GB x 2) |
| Type | DDR5 |
| Speed (MT/s) | 5600 |
| CAS latency | 46 |
| Voltage | 1.1 V |
| Profile | XMP 3.0 and EXPO: 5600 MT/s 46-45-45; 5200 MT/s 42-42-42 |
| Modules | 2 x 64 GB, 2Rx8 |
| Warranty | Limited lifetime (10 years in Germany and France) |
Pros
- Runs at the standard 1.1 V
- XMP 3.0 and EXPO profiles listed
- Crucial lists a limited lifetime warranty, with 10 years in Germany and France
- Two dual-rank 2Rx8 modules, matching the DDR5-5600 AMD lists for two dual-rank DIMMs
Cons
- CL46 is the loosest timing of the three kits
- The lookalike non-Pro kit lacks profiles, so the part number matters
Best for builders who prefer stock voltage and a 128 GB kit with both profiles
Not for buyers who want the highest speed from a 128 GB kit
Sources: Maker: crucial.in (opens in a new tab), Benchmark: dev.to (opens in a new tab)
G.Skill Ripjaws S5 96GB (2x48GB) DDR5-6400 CL32
G.Skill- Capacity96 GB
- TypeDDR5
- SpeedUp to 6400
- CAS32-39-39-102
Not every local LLM user needs 128 GB. Ninety-six gigabytes holds a 70B model at Q6_K, a 58 GB file in the Ollama library, and the 65 GB gpt-oss 120B model with space for context. It has less capacity than the other two kits and a higher rated speed.
G.Skill tests the kit at up to DDR5-6400 with 32-39-39-102 timings at 1.35 V. The SPD default is 5600 MT/s at 1.10 V, and the kit boots there until you enable XMP. G.Skill lists Intel XMP 3.0 only, describes the series as for compatible Intel platforms and gives a low module height of 33 mm.
Intel gives DDR5-6400 as the top speed for the Core Ultra 9 285K, so this kit's rating sits at the platform's stated limit. On a Ryzen board you would have no EXPO profile to use, and AMD caps a pair of DIMMs at DDR5-5600 on the 9950X.
| Spec | Detail |
|---|---|
| Capacity (kit) | 96 GB (48 GB x 2) |
| Type | DDR5 |
| Speed (MT/s) | Up to 6400 (tested XMP) |
| CAS latency | 32-39-39-102 |
| Voltage | 1.35 V (XMP); SPD 1.10 V |
| Profile | Intel XMP 3.0 |
| Module type | Unbuffered, non-ECC |
| Warranty | Limited lifetime |
Pros
- Fastest rated speed of the three kits
- Low 33 mm module height
- Fits a 70B model at Q6_K plus context
- Rating matches Intel's listed DDR5-6400 limit
Cons
- Intel XMP only
- 96 GB cannot hold a 70B model at Q8_0
Best for Intel builders who run 70B models at 5-bit or 6-bit in RAM
Not for buyers on Ryzen boards and anyone who wants 70B at 8-bit
Sources: Maker: gskill.com (opens in a new tab), Benchmark: dev.to (opens in a new tab)
Performance and benchmarks#
Large-capacity kits have almost no direct benchmark coverage, so the results below describe the mechanisms that matter. Maxim Saplin ran CPU-only tests with a four-stick 96 GB setup and a two-stick 64 GB setup on a Core i5-13600KF. This is community-reported data.
Mistral 7B Q6_K CPU-only generation: Token generation speed (tokens/s, higher is better)
2 products. Highest: 2 DIMMs (64 GB), 9.66 tokens/s. Lowest: 4 DIMMs (96 GB), 9.42 tokens/s.
| Product | tokens/s |
|---|---|
| 2 DIMMs (64 GB) | 9.66 |
| 4 DIMMs (96 GB) | 9.42 |
About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.
At DDR5-4800 on both layouts, Mistral 7B ran at 9.66 tokens/s with two sticks and 9.42 with four. The four-stick kit could not hold its 6000 profile. That is the argument for two big modules: the 2x64 GB and 2x48 GB kits in this guide avoid the problem.
Llama 3.2 3B Q4_K_M, llama.cpp tg128: Token generation speed (tokens/s, higher is better)
2 products. Highest: Dual channel, 11.59 tokens/s. Lowest: Single channel, 7.36 tokens/s.
| Product | tokens/s |
|---|---|
| Dual channel | 11.59 |
| Single channel | 7.36 |
About this data: figures are copied from the named source, not measured by us, and the test system and settings are those the source states above. See all benchmark sources and credits.
Channel count matters too. InsiderLLM measured Llama 3.2 3B at Q4_K_M with llama.cpp on an older DDR4 mini PC and found 7.36 tokens/s on a single stick and 11.59 on a pair. The hardware is old, so read the result as a demonstration of the effect, not as a prediction for DDR5. Phoronix reported the same direction on a Core Ultra 7 270K Plus, where llama.cpp ran significantly faster with two DDR5-6400 sticks than with one.
Puget Systems measured the other side of the story. On Threadripper PRO, DDR5 speed made little difference to llama.cpp, while dropping from eight channels to four cost about 14 percent on the 9995WX. Total bandwidth is what counts. A desktop with two channels has a fixed ceiling, so capacity kits cannot change it.
Capacity is the other half of the picture, and Corsair's own article on DRAM for AI is the clearest published guide. Corsair reports that 32 GB and 48 GB configurations could not fit a 70B DeepSeek-R1 distill when it ran on the CPU alone, and that 32 GB is on the edge even with a GeForce RTX 4090 in the system. Its advice is 64 GB as a floor and 96 GB for an AI-only machine that should last. The 128 GB kits here sit one step beyond that advice, which suits buyers who want to load a high-precision 70B model or keep two models resident at once.
Be realistic about speed. A model larger than the memory of your graphics card will run at the pace of system RAM. Gaessler's notes say about five threads are enough to saturate dual-channel desktop memory, so a processor with many cores gains little. The practical gain of a 128 GB kit is that a model runs at all, not that it runs quickly.
Value for money#
With a fixed ceiling on bandwidth, value in this guide is capacity per slot and platform fit.
- 128 GB tier (TeamGroup T-Create Expert 128GB): the highest rated speed of the two 128 GB kits, with both profiles at 1.25 V.
- 128 GB tier, stock voltage (Crucial Pro 128GB): the lowest voltage of the three, at looser timings.
- 96 GB tier (G.Skill Ripjaws S5 96GB): the smallest kit that still runs a 65 GB model, and the fastest rated.
Think about the board. A 128 GB kit of 64 GB modules needs a BIOS that supports them, and not every older AM5 or Intel board does. Prices change, so check current prices before buying.
Weigh a GPU upgrade against more memory as well. A card with more VRAM keeps a model out of system RAM and avoids the bandwidth penalty. Our guide to VRAM for LLMs explains the sizes.
How to choose#
- Read the model's file size first. In the Ollama library, Llama 3.3 70B is 58 GB at Q6_K, 75 GB at Q8_0 and 141 GB at FP16. The gpt-oss 120B model is 65 GB.
- Leave a margin. Corsair says 32 GB and 48 GB could not fit a 70B DeepSeek-R1 distill running CPU-only, and it recommends 64 GB at minimum and 96 GB for future-proofing.
- Know the limits of 128 GB. Qwen3 235B at Q4_K_M is 142 GB, so it will not fit.
- Use two modules. Four-DIMM layouts bring a much lower supported speed on the Ryzen 9 9950X.
- Match the profile to the board. TeamGroup and Crucial list both XMP and EXPO. G.Skill's kit lists XMP only.
- Check the board's memory support list. Makers say the final speed depends on the board and processor.
- Decide between size and speed. DDR5 speed versus capacity for AI covers the trade-off, and the sizing guide for local models lists file sizes.
Common mistakes#
- Counting on 128 GB for any model. The largest open models still exceed it.
- Using four modules for capacity. Saplin's four-stick setup dropped to 4800, and AMD's listed speed falls for four DIMMs.
- Skipping the part number. Crucial sells a similar 128 GB kit with no profiles.
- Expecting VRAM-class speed. Models in system RAM run at the speed of system memory bandwidth.
- Mixing kits. G.Skill and TeamGroup advise against mixing.
Verdict#
Best overall TeamGroup T-Create Expert 128GB (2x64GB) DDR5-6000 CL40See the details
TeamGroup T-Create Expert 128GB (2x64GB) DDR5-6000 CL40 is the best large kit for local LLMs, because it combines 128 GB in two modules with a DDR5-6000 rating and profiles for both AMD and Intel. Choose the Crucial Pro 128GB if you want the standard 1.1 V and a lifetime warranty. Choose the G.Skill Ripjaws S5 96GB if you are on Intel and your largest model fits in 96 GB.
If you are still deciding on size, start with the best RAM for AI and machine learning or the best 64GB DDR5 RAM kit. The RAM and memory page lists all guides.
Frequently asked questions#
Is 128 GB of RAM worth it for local LLMs?
It is worth it if you run 70B models at 8-bit, which is a 75 GB file in the Ollama library. For smaller models, 64 GB is enough.
Can I run a 235B model in 128 GB?
No. The Qwen3 235B model at Q4_K_M is 142 GB in the Ollama library, which is above 128 GB before the system and context are counted. A smaller model or heavier compression is needed for that capacity.
Is CPU inference fast enough to be useful?
It runs slower than a GPU. Gaessler's notes say CPU generation is limited by memory bandwidth, so a large model in RAM generates tokens slowly.
Should I buy 2x64 GB or 4x32 GB?
Two modules. A four-DIMM layout lowers the supported speed on AMD's Ryzen 9 9950X. A pair of 64 GB sticks reaches the same total at a faster supported speed, and leaves two slots free for later changes.
Does 128 GB work on any motherboard?
Not always. AMD and Intel list 256 GB as the platform maximum on current chips, but check your board maker's memory support list for 64 GB modules.
Do these kits support ECC?
TeamGroup lists on-die ECC for the series, which is built into DDR5. G.Skill lists its modules as non-ECC. On-die ECC is not the same as full error correction, which needs supported workstation boards and memory.
Is the Crucial non-Pro 128 GB kit the same?
No. Crucial's page for that kit lists no XMP or EXPO profile, so it runs on JEDEC settings. That means a lower default speed, so check that you are happy with it before choosing the cheaper kit.
Is 96 GB enough instead of 128 GB?
For a 70B model at Q6_K, which is 58 GB, or gpt-oss 120B at 65 GB, yes. A 70B model at Q8_0 is 75 GB and fits better in 128 GB.
Sources & credits#
We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name.
- TeamGroup, T-Create Expert DDR5 128GB product page (opens in a new tab)
- Crucial India, Pro DDR5 128GB kit CP2K64G56C46U5 (opens in a new tab)
- Crucial India, DDR5 128GB kit CT2K64G56C46U5 (opens in a new tab)
- G.Skill, Ripjaws S5 96GB DDR5-6400 specification (opens in a new tab)
- AMD, Ryzen 9 9950X product page (opens in a new tab)
- Intel, Core Ultra 9 285K specifications (opens in a new tab)
- Maxim Saplin, DDR5 speed and LLM inference (opens in a new tab)
- InsiderLLM, dual-channel RAM for local LLM CPU inference (opens in a new tab)
- Puget Systems, impact of RAM speed on Threadripper PRO (opens in a new tab)
- Phoronix, single versus dual channel memory on Core Ultra 7 270K Plus (opens in a new tab)
- Corsair Explorer, choosing the best DRAM for AI (opens in a new tab)
- llama.cpp performance notes, Johannes Gaessler (opens in a new tab)
- Ollama library, Llama 3.3 tags (opens in a new tab)
- Ollama library, gpt-oss tags (opens in a new tab)
- Ollama library, Qwen3 tags (opens in a new tab)
Facts and figures in this article come from these sources. Articles are compiled from the makers' published specifications and the benchmarks named above, and checked by the NextAIGrid Editorial Team.
Disclaimer
This guide is for information only. Specifications and availability change, so check the maker's page. Found an error? Tell us.
How was this buying guide put together?
It draws on 4 named sources: maker specification pages and published third-party benchmarks, listed in the article. We have not tested these products ourselves. Specs come from the makers; benchmark figures come from the sources we name. Specifications and sources were last checked on 17 November 2026. Warranty terms and availability differ by brand and seller in India, so check the maker's India page for the warranty period and buy from an authorised seller. Read how we research and our editorial policy.
Spotted an error? Tell us. Confirmed fixes are listed on the corrections page.