Local LLM Hardware Picker
约 135 字
Local LLM Hardware Picker
Pick your GPU, RAM or Mac to see which models it can run, which quantization to use and how fast it will be. Or start from a model and see what hardware it needs.
Context is how much the model can keep in mind at once. One token is roughly ¾ of an English word. For coding or agents, use 64K or more.
125B main model plus a 51B N-gram table, with only 6B active per token. A 16–24GB GPU with 64GB+ of RAM can run it; users mostly report 60–130 tok/s.
Fits in VRAM 9, fastest, pick these first
Quality: slight loss
Engine: llama.cpp
Only 3B active, so it’s very fast, even on 12–16GB GPUs.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 72.3GB | VRAM + RAM | 14–22 tok/s no MTP 11–15 | near lossless |
| Q8_0 | 39.5GB | VRAM + RAM | 30–47 tok/s no MTP 25–32 | near lossless |
| Q6_K | 31.2GB | VRAM + RAM | 44–66 tok/s no MTP 37–45 | near lossless |
| Q5_K_M | 27.3GB | VRAM + RAM | 56–82 tok/s no MTP 47–56 | slight loss |
| Q4_K_M | 23.6GB | VRAM + RAM | 75–105 tok/s no MTP 63–72 | slight loss |
| IQ4_XS / MXFP4 | 21.2GB | Fits in VRAM | 93–125 tok/s no MTP 78–86 | slight loss |
| Q3_K_M | 19.4GB | Fits in VRAM | 95–127 tok/s no MTP 80–87 | noticeable loss |
| IQ3_S | 17.7GB | Fits in VRAM | 97–130 tok/s no MTP 82–89 | clear loss |
| IQ2_XS | 12.8GB | Fits in VRAM | 104–137 tok/s no MTP 87–94 | heavy loss |
Quality: slight loss
Engine: llama.cpp
A MoE with 3.8B active. Fast.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 52.2GB | VRAM + RAM | 13–20 tok/s no MTP 11–14 | near lossless |
| Q8_0 | 28.6GB | VRAM + RAM | 36–55 tok/s no MTP 30–38 | near lossless |
| Q6_K | 22.6GB | VRAM + RAM | 68–96 tok/s no MTP 57–66 | near lossless |
| Q5_K_M | 19.7GB | Fits in VRAM | 83–113 tok/s no MTP 69–78 | slight loss |
| Q4_K_M | 17.1GB | Fits in VRAM | 87–119 tok/s no MTP 73–81 | slight loss |
| IQ4_XS / MXFP4 | 15.3GB | Fits in VRAM | 90–122 tok/s no MTP 76–84 | slight loss |
| Q3_K_M | 14.1GB | Fits in VRAM | 93–125 tok/s no MTP 78–86 | noticeable loss |
| IQ3_S | 12.8GB | Fits in VRAM | 96–128 tok/s no MTP 80–88 | clear loss |
| IQ2_XS | 9.3GB | Fits in VRAM | 104–136 tok/s no MTP 87–93 | heavy loss |
Quality: slight loss
Engine: llama.cpp
A mid-small model that also accepts audio input.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 25.6GB | VRAM + RAM | 13–26 tok/s no MTP 9–12 | near lossless |
| Q8_0 | 14.4GB | Fits in VRAM | 50–94 tok/s no MTP 34–44 | near lossless |
| Q6_K | 11.5GB | Fits in VRAM | 62–114 tok/s no MTP 42–53 | near lossless |
| Q5_K_M | 10.2GB | Fits in VRAM | 69–127 tok/s no MTP 47–59 | slight loss |
| Q4_K_M | 8.9GB | Fits in VRAM | 78–141 tok/s no MTP 53–66 | slight loss |
| IQ4_XS / MXFP4 | 8.1GB | Fits in VRAM | 85–153 tok/s no MTP 57–71 | slight loss |
| Q3_K_M | 7.5GB | Fits in VRAM | 91–162 tok/s no MTP 61–76 | noticeable loss |
| IQ3_S | 6.9GB | Fits in VRAM | 97–173 tok/s no MTP 66–81 | clear loss |
| IQ2_XS | 5.3GB | Fits in VRAM | 122–213 tok/s no MTP 83–99 | heavy loss |
Quality: near lossless
Engine: llama.cpp
A small model that runs on laptops and older GPUs.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 11.1GB | Fits in VRAM | 46–58 tok/s | near lossless |
| Q8_0 | 7.4GB | Fits in VRAM | 69–84 tok/s | near lossless |
| Q6_K | 6.4GB | Fits in VRAM | 78–95 tok/s | near lossless |
| Q5_K_M | 6GB | Fits in VRAM | 84–101 tok/s | slight loss |
| Q4_K_M | 5.6GB | Fits in VRAM | 90–107 tok/s | slight loss |
| IQ4_XS / MXFP4 | 5.3GB | Fits in VRAM | 94–111 tok/s | slight loss |
| Q3_K_M | 5.1GB | Fits in VRAM | 98–115 tok/s | noticeable loss |
| IQ3_S | 4.9GB | Fits in VRAM | 102–119 tok/s | clear loss |
| IQ2_XS | 4.3GB | Fits in VRAM | 114–131 tok/s | heavy loss |
Quality: near lossless
Engine: llama.cpp
Native 4-bit at about 13GB. Runs easily on a 16GB GPU.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| MXFP4 (native) | 15.5GB | Fits in VRAM | 70–79 tok/s | near lossless |
Quality: slight loss
Engine: llama.cpp
The first pick for 8GB GPUs.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 21.1GB | Fits in VRAM | 25–32 tok/s | near lossless |
| Q8_0 | 12.7GB | Fits in VRAM | 41–52 tok/s | near lossless |
| Q6_K | 10.6GB | Fits in VRAM | 49–61 tok/s | near lossless |
| Q5_K_M | 9.6GB | Fits in VRAM | 54–67 tok/s | slight loss |
| Q4_K_M | 8.6GB | Fits in VRAM | 59–74 tok/s | slight loss |
| IQ4_XS / MXFP4 | 8GB | Fits in VRAM | 64–79 tok/s | slight loss |
| Q3_K_M | 7.5GB | Fits in VRAM | 67–83 tok/s | noticeable loss |
| IQ3_S | 7.1GB | Fits in VRAM | 71–87 tok/s | clear loss |
| IQ2_XS | 5.8GB | Fits in VRAM | 86–103 tok/s | heavy loss |
Quality: slight loss
Engine: llama.cpp
Google’s dense model. Good at many languages and ships with a draft model for speculative decoding.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 65.5GB | VRAM + RAM | 2–4 tok/s no MTP 1–2 | near lossless |
| Q8_0 | 36.7GB | VRAM + RAM | 5–10 tok/s no MTP 3–5 | near lossless |
| Q6_K | 29.4GB | VRAM + RAM | 8–16 tok/s no MTP 6–8 | near lossless |
| Q5_K_M | 26GB | VRAM + RAM | 13–25 tok/s no MTP 9–12 | slight loss |
| Q4_K_M | 22.7GB | VRAM + RAM | 24–47 tok/s no MTP 16–22 | slight loss |
| IQ4_XS / MXFP4 | 20.6GB | Fits in VRAM | 38–72 tok/s no MTP 26–33 | slight loss |
| Q3_K_M | 19.1GB | Fits in VRAM | 41–77 tok/s no MTP 28–36 | noticeable loss |
| IQ3_S | 17.5GB | Fits in VRAM | 44–84 tok/s no MTP 30–39 | clear loss |
| IQ2_XS | 13.3GB | Fits in VRAM | 59–108 tok/s no MTP 40–51 | heavy loss |
Quality: slight loss
Engine: llama.cpp
Dense 27B and one of the most popular local models. 24GB of VRAM runs Q4; on 32GB we suggest Q4 to leave room for long context, though Q5–Q6 also fit.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 59.3GB | VRAM + RAM | 2–4 tok/s no MTP 1–2 | near lossless |
| Q8_0 | 34GB | VRAM + RAM | 6–11 tok/s no MTP 4–5 | near lossless |
| Q6_K | 27.6GB | VRAM + RAM | 10–20 tok/s no MTP 7–9 | near lossless |
| Q5_K_M | 24.5GB | VRAM + RAM | 16–32 tok/s no MTP 11–15 | slight loss |
| Q4_K_M | 21.7GB | Fits in VRAM | 37–71 tok/s no MTP 25–33 | slight loss |
| IQ4_XS / MXFP4 | 19.8GB | Fits in VRAM | 41–77 tok/s no MTP 28–36 | slight loss |
| Q3_K_M | 18.5GB | Fits in VRAM | 44–83 tok/s no MTP 30–39 | noticeable loss |
| IQ3_S | 17.1GB | Fits in VRAM | 48–90 tok/s no MTP 32–42 | clear loss |
| IQ2_XS | 13.4GB | Fits in VRAM | 62–114 tok/s no MTP 42–53 | heavy loss |
Quality: slight loss
Engine: llama.cpp
The previous 27B, with the most guides and tuning experience available.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 59.3GB | VRAM + RAM | 2–4 tok/s no MTP 1–2 | near lossless |
| Q8_0 | 34GB | VRAM + RAM | 6–11 tok/s no MTP 4–5 | near lossless |
| Q6_K | 27.6GB | VRAM + RAM | 10–20 tok/s no MTP 7–9 | near lossless |
| Q5_K_M | 24.5GB | VRAM + RAM | 16–32 tok/s no MTP 11–15 | slight loss |
| Q4_K_M | 21.7GB | Fits in VRAM | 37–71 tok/s no MTP 25–33 | slight loss |
| IQ4_XS / MXFP4 | 19.8GB | Fits in VRAM | 41–77 tok/s no MTP 28–36 | slight loss |
| Q3_K_M | 18.5GB | Fits in VRAM | 44–83 tok/s no MTP 30–39 | noticeable loss |
| IQ3_S | 17.1GB | Fits in VRAM | 48–90 tok/s no MTP 32–42 | clear loss |
| IQ2_XS | 13.4GB | Fits in VRAM | 62–114 tok/s no MTP 42–53 | heavy loss |
VRAM + RAM 4, runs, but part sits in RAM, so speed depends on RAM bandwidth
Quality: slight loss
Engine: Strata
125B main model plus a 51B N-gram table, with only 6B active per token. A 16–24GB GPU with 64GB+ of RAM can run it; users mostly report 60–130 tok/s.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality | Official tests |
|---|---|---|---|---|---|
| BF16 | 254GB +103GB | Needs about 18GB more RAM | — | near lossless | Top-1 100% |
| Q8_0 | 136GB +54.5GB | VRAM + RAM | 26–44 tok/s no MTP 19–25 | near lossless | Top-1 94.1% |
| UD-Q6_K_XL | 117GB +54.5GB | VRAM + RAM | 31–52 tok/s no MTP 23–30 | near lossless | Top-1 94.1% |
| UD-Q5_K_XL | 106GB +54.5GB | VRAM + RAM | 35–59 tok/s no MTP 26–33 | near lossless | Top-1 93.7% |
| UD-Q4_K_XL | 81.7GB +32.3GB | VRAM + RAM | 48–80 tok/s no MTP 36–46 | slight loss | Top-1 92.3% |
| UD-IQ4_XS | 67.6GB +28.7GB | VRAM + RAM | 59–97 tok/s no MTP 45–55 | slight loss | Top-1 89.6% |
| UD-Q3_K_XL | 63.4GB +28.7GB | VRAM + RAM | 63–103 tok/s no MTP 48–59 | noticeable loss | Top-1 88.3% |
| UD-IQ3_XXS | 57.5GB +28.7GB | VRAM + RAM | 70–113 tok/s no MTP 53–64 | clear loss | Top-1 85.4% |
| UD-Q2_K_XL | 53.5GB +28.7GB | VRAM + RAM | 75–121 tok/s no MTP 57–69 | clear loss | Top-1 82.7% |
| UD-IQ1_M | 41.9GB +28.7GB | VRAM + RAM | 95–148 tok/s no MTP 72–84 | heavy loss | Top-1 79.7% |
| UD-IQ1_S | 40.3GB +28.7GB | VRAM + RAM | 99–153 tok/s no MTP 75–87 | heavy loss | Top-1 77.3% |
| GSQ-RCO IQ3_S | 57.5GB +28.7GB | VRAM + RAM | 70–113 tok/s no MTP 53–64 | slight loss | task avg 93.26 (100.2% of original) |
| GSQ-RCO IQ3_XXS | 49.7GB +28.7GB | VRAM + RAM | 81–129 tok/s no MTP 61–73 | slight loss | task avg 92.57 (99.4% of original) |
| GSQ-RCO IQ2_XS | 41.9GB +28.7GB | VRAM + RAM | 95–148 tok/s no MTP 72–84 | clear loss | task avg 89.16 (95.7% of original) |
| GSQ-RCO Q2_0 | 40.3GB +28.7GB | VRAM + RAM | 99–153 tok/s no MTP 75–87 | clear loss | task avg 89.07 (95.7% of original) |
Quality: near lossless
Engine: llama.cpp
Native MXFP4 at about 63GB. A good fit for 128GB Macs or AI Max machines.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| MXFP4 (native) | 66.3GB | VRAM + RAM | 22–28 tok/s | near lossless |
Quality: noticeable loss
Engine: Maya
321B total, 18B active. Strong at coding and agent tasks.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 645GB | Needs about 253GB more RAM | — | near lossless |
| Q8_0 | 344GB | Needs about 72GB more RAM | — | near lossless |
| Q6_K | 268GB | Needs about 26GB more RAM | — | near lossless |
| Q5_K_M | 232GB | Needs about 5GB more RAM | — | near lossless |
| Q4_K_M | 198GB | VRAM + RAM | 6–8 tok/s no MTP 6–8 | slight loss |
| IQ4_XS / MXFP4 | 176GB | VRAM + RAM | 8–12 tok/s no MTP 8–11 | slight loss |
| Q3_K_M | 160GB | VRAM + RAM | 12–16 tok/s no MTP 11–15 | noticeable loss |
| IQ3_S | 143GB | VRAM + RAM | 19–26 tok/s no MTP 18–24 | noticeable loss |
| IQ2_XS | 99.3GB | VRAM + RAM | 29–40 tok/s no MTP 28–36 | heavy loss |
Quality: heavy loss
Engine: Maya
552B total, 16B active while generating. Its KV cache is tiny, so long context costs little memory. A separate 196B Engram table can live on SSD. If it doesn’t fit in VRAM, use Project Maya (Strata dropped it after 0.2.13); llama.cpp supports it too.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 1105GB +392GB | Needs about 529GB more RAM | — | near lossless |
| Q8_0 | 588GB +208GB | Needs about 218GB more RAM | — | near lossless |
| Q6_K | 457GB +162GB | Needs about 140GB more RAM | — | near lossless |
| Q5_K_M | 395GB +140GB | Needs about 102GB more RAM | — | near lossless |
| Q4_K_M | 336GB +119GB | Needs about 67GB more RAM | — | slight loss |
| IQ4_XS / MXFP4 | 298GB +105GB | Needs about 44GB more RAM | — | slight loss |
| Q3_K_M | 270GB +95.6GB | Needs about 28GB more RAM | — | noticeable loss |
| IQ3_S | 243GB +85.8GB | Needs about 11GB more RAM | — | noticeable loss |
| IQ2_XS | 167GB +58.8GB | VRAM + RAM | 14–19 tok/s no MTP 13–17 | heavy loss |
Won't run 1, not enough memory, or the engine doesn’t support it
One of the strongest open models today. At 2.8T parameters it needs a server cluster; personal machines can’t run it.
| Quant | Needs | Runs? (128K context) | Est. speed | Quality |
|---|---|---|---|---|
| BF16 | 5605GB | Needs about 5462GB more RAM | — | near lossless |
| Q8_0 | 2980GB | Needs about 2837GB more RAM | — | near lossless |
| Q6_K | 2315GB | Needs about 2172GB more RAM | — | near lossless |
| Q5_K_M | 2000GB | Needs about 1857GB more RAM | — | near lossless |
| Q4_K_M | 1703GB | Needs about 1559GB more RAM | — | slight loss |
| IQ4_XS / MXFP4 | 1510GB | Needs about 1367GB more RAM | — | slight loss |
| Q3_K_M | 1370GB | Needs about 1227GB more RAM | — | noticeable loss |
| IQ3_S | 1230GB | Needs about 1087GB more RAM | — | noticeable loss |
| IQ2_XS | 845GB | Needs about 702GB more RAM | — | heavy loss |
Fits in VRAM 40, fastest
No measurements for this setup yet.
No measurements for this setup yet.
Engine: MLX
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
Engine: MLX
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
Engine: MLX
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
Engine: MLX
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
Engine: vLLM/SGLang
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
VRAM + RAM 38, assuming 128GB of RAM
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
No measurements for this setup yet.
Multiple machines linked together (user-reported)
Several DGX Sparks, Macs or AI Max boxes linked over Ethernet or Thunderbolt. The list above covers single machines only, so these setups are listed here.
How much quantization costs
Quantization shrinks a model. The harder you squeeze, the less memory it needs and the faster it runs, but answer quality drops. On the left is a model with official per-quant tests; on the right, rules of thumb for common quants.
Qwen3.8-Flash-Next: Top-1 agreement with the original BF16 (Unsloth official tests)
Top-1 agreement is how often the quant picks the same next token as the original. It measures closeness to the original, not a proportional drop in task scores; even Q8 differs from the original by about 6%, which shows how sensitive this metric is to tiny numeric differences (bars start at 70% to make differences easier to compare). File sizes are the current Hugging Face files (re-uploaded 2026-11-04) and include the N-gram table, which can sit in RAM or on SSD; Top-1 comes from Unsloth’s tests of the earlier upload.
ISTA-DASLab GSQ-RCO (what Strata ships)
No Top-1 numbers are published for these, only task results (ISTA-DASLab model card; AIME25 marked * is back-calculated from the published three-task average). Both 3-bit builds lose essentially nothing; IQ3_S edging past the original is benchmark noise. The model card calls the smaller IQ3_XXS the best operating point; this page defaults to the top-scoring IQ3_S, and IQ3_XXS is the pick when memory is tight. The two 2-bit builds score about 4 points lower in exchange for a smaller size, and Q2_0 generates faster and more steadily. Task scores and the Top-1 figures above measure different things and can’t be compared directly.
Rules of thumb for common quants
The same quant name can perform quite differently depending on who made it (Unsloth, bartowski, GSQ-RCO and others). Very large MoE models tolerate low quants better; for dense models like 27B, stay at Q4 or above.
Choosing an inference engine
Easiest to start with, works on almost any GPU, and can spill into system RAM. Ollama and LM Studio are built on it.
Multiple GPUs add memory but not speed; throughput with many users is modest.
Tensor parallelism adds up bandwidth across GPUs, the best throughput for many users, and solid MTP speculative decoding.
The model must fit entirely in VRAM. Setup is complex, some AMD cards need patched builds, and Intel GPUs use the XPU backend with fewer supported models.
Apple’s own framework. Usually faster than llama.cpp on a Mac, with good long-context caching.
Mac only.
Built for large MoE models: a 16–24GB GPU plus lots of RAM can run 125B-class models fast for a single user. It ships ISTA-DASLab’s GSQ-RCO quants (Q2_0 to IQ3_S) out of the box.
Best for one conversation at a time; extra sessions queue. Supports only a few Qwen models and changes quickly (dropped DeepSeek V4.1 Flash after 0.2.13). There is a Metal build for Macs, but it is slower than MLX.
Open source and built for big MoE models: hot experts stay in VRAM, the next tier in RAM, the rest streams from SSD and is rebalanced as you chat. Runs on 1–16 NVIDIA GPUs with OpenAI- and Anthropic-compatible APIs.
Only GLM-5.3-Flash and DeepSeek V4.1 Flash for now. Linux first, AMD is experimental, and it needs a fast NVMe SSD.
How to read the numbers
- Speed is how fast the model writes its answer, in tokens per second. The main figure is estimated at the context length you picked, with MTP speculative decoding on for models that support it (nearly everyone running locally turns it on); the speed without MTP is shown below for comparison. The range reflects efficiency differences between engine versions, drivers and cooling; tasks where MTP hits often, like coding or fixed-format tool calls, land nearer the top.
- Where users have reported real numbers for a setup and the quant (below Q6), engine, context (128K or less), whether the model fits fully in VRAM, and memory size all match what they tested, those numbers are shown (reports combined from several users first; a single user’s report is used only when it roughly agrees with the estimate). Otherwise you see an estimate from memory bandwidth and model size, calibrated against those reports, with the reports listed below for reference. The same hardware can vary a lot with engine version and settings.
- How the user-reported range is built: multiple posts by the same user are merged into one median first (so one person cannot skew a group), then we take the middle half across independent users, which trims before/after-tuning numbers and unusual builds. Prompt-processing speeds, numbers above what memory bandwidth physically allows, reports whose engine doesn’t match the hardware, numbers quoted in other people’s replies, reports more than 2× away from the rest and groups that contradict matching hand-checked reports are all left out; groups that include custom unmodeled engines are shown for reference only.
- When a model doesn’t fit in VRAM, part of it goes into system RAM. It still runs, but slower, and how much slower depends on RAM bandwidth: quad- and 8-channel memory is much faster. Engines like Strata and Maya also keep rarely used experts and the N-gram table on SSD, so they need much less RAM than the file size.
- Lower quantization saves memory and runs faster at some cost in quality. Q4_K_M is the usual sweet spot; very large MoE models tolerate low quantization better.
- Intelligence stars draw on the Artificial Analysis Intelligence Index and official benchmarks, summarized by this site. Results vary by task.
- Memory figures assume an 8-bit KV cache (with unified K=V global layers accounted for on Gemma 4). KV sizes for some new models are estimated from their architecture. The N-gram table in Qwen3.8-Flash-Next and the Engram table in DeepSeek V4.1 Flash are lookup parameters: llama.cpp, Strata and Maya can keep them in RAM or on SSD, so they don’t count toward what must fit in VRAM; MLX and standard vLLM setups are modeled as loading them in full (newer vLLM builds can also offload N-gram tables to host RAM).
- github.com/mw00/project-maya
- unsloth.ai/docs/models/qwen3.8-next
- huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-GGUF
- github.com/Niko1221/Strata
- huggingface.co
- artificialanalysis.ai
- github.com/ggml-org/llama.cpp
- github.com/vllm-project/vllm
- github.com/sgl-project/sglang
- github.com/ml-explore/mlx
- lcz.me
- nvidia.com
- amd.com
- intel.com
- apple.com
