mistralai · text

Devstral-2-123B-Instruct-2512

mistralai/Devstral-2-123B-Instruct-2512

Devstral-2-123B-Instruct-2512 at Q4_K_M is exactly 74,897,652,896 bytes (69.75 GiB / 74.90 GB) — an effective 4.793 bits per weight, not the nominal 4. Its KV cache at 32K is 11.00 GiB.

From the file· summed from 2 file(s)From the file· KV per layer
Parameters
125B
Architecture
llama
88 layers
Context
262,144
native (config.json)
License
other

Shipped quantizations

● exact bytes, summed from published files
QuantSize●Exact bytes●Effective bpw●Tensors●Publisher
IQ1_S25.33 GiB27,194,260,8321.740—bartowski
UD-IQ1_S26.49 GiB28,445,876,4481.820—unsloth
IQ1_M27.59 GiB29,620,796,7681.895—bartowski
UD-IQ1_M28.47 GiB30,565,310,6881.956—unsloth
IQ2_XXS31.35 GiB33,665,023,3282.154—bartowski
UD-IQ2_XXS32.04 GiB34,404,671,7122.201—unsloth
IQ2_XS34.75 GiB37,315,640,6722.388—bartowski
IQ2_S37.01 GiB39,741,390,1762.543—bartowski
IQ2_M40.03 GiB42,976,771,4242.750—bartowski
UD-IQ2_M40.55 GiB43,543,404,7682.786—unsloth
Q2_K43.39 GiB46,591,212,8962.981—bartowski
Q2_K43.39 GiB46,591,221,9842.981—unsloth
Q2_K_L43.74 GiB46,968,709,3443.005—unsloth
Q2_K_L44.86 GiB48,164,076,8963.082—bartowski
IQ3_XXS45.04 GiB48,366,189,9203.095—bartowski
UD-IQ3_XXS45.60 GiB48,963,887,3283.133—unsloth
IQ3_XS2 shards48.11 GiB51,659,767,3603.305—bartowski
Q3_K_S2 shards50.63 GiB54,367,452,7043.479—bartowski
Q3_K_S2 shards50.63 GiB54,367,461,7923.479—unsloth
IQ3_M2 shards52.89 GiB56,793,988,6723.634—bartowski
Q3_K_M2 shards56.46 GiB60,620,373,5683.879—bartowski
Q3_K_M2 shards56.46 GiB60,620,382,6563.879—unsloth
Q3_K_L2 shards61.53 GiB66,071,919,7764.228—lmstudio-community
Q3_K_L2 shards61.53 GiB66,071,920,1604.228—bartowski
IQ4_XS2 shards62.47 GiB67,074,620,9604.292—bartowski
IQ4_XS2 shards62.51 GiB67,124,961,6964.295—unsloth
IQ4_NL2 shards66.03 GiB70,896,680,4804.536—bartowski
IQ4_NL2 shards66.03 GiB70,896,689,5684.536—unsloth
Q4_02 shards66.12 GiB71,000,489,5364.543—bartowski
Q4_02 shards66.12 GiB71,000,498,5924.543—unsloth
Q4_K_S2 shards66.36 GiB71,249,002,0164.559—bartowski
Q4_K_S2 shards66.36 GiB71,249,011,1044.559—unsloth
Q4_K_M2 shards69.75 GiB74,897,652,8964.793—lmstudio-community
Q4_K_M2 shards69.75 GiB74,897,653,3124.793—bartowski
Q4_K_M2 shards69.75 GiB74,897,662,4004.793—unsloth
Q4_K_L2 shards70.87 GiB76,093,029,9204.869—bartowski
Q4_12 shards73.08 GiB78,471,593,5045.021—bartowski
Q4_12 shards73.08 GiB78,471,602,5925.021—unsloth
Q5_K_S3 shards80.27 GiB86,184,918,6885.515—bartowski
Q5_K_S2 shards80.27 GiB86,184,927,6805.515—unsloth

KV cache by context

computed per layer
ContextKV cache (f16)●Flat formulaOverstated byFull / windowed / recurrent
4,0961.38 GiB1.38 GiB—88 / 0 / 0
8,1922.75 GiB2.75 GiB—88 / 0 / 0
16,3845.50 GiB5.50 GiB—88 / 0 / 0
32,76811.00 GiB11.00 GiB—88 / 0 / 0
65,53622.00 GiB22.00 GiB—88 / 0 / 0
131,07244.00 GiB44.00 GiB—88 / 0 / 0

Compare with

same modality, comparable size
vs Qwen3-Coder-30B-A3B-Instructvs Qwen3.6-27Bvs Qwen3.8-27Bvs gemma-4-E4B-itvs Qwen3-30B-A3B-Thinking-2507vs Qwen3-8BAll alternatives →Will it fit on my card? →

Will it run on your card?

full quant x context sweep
By VRAM:8GB12GB16GB24GB32GB48GB
Radeon RX 6500 XT 4GBGeForce RTX 3050 6GBGeForce RTX 5050 8GBGeForce RTX 3080 10GBGeForce RTX 2080 Ti 11GBGeForce RTX 5070 12GBGeForce RTX 5060 Ti 16GBApple M3 Pro 18GBGeForce RTX 3080 Ti 20GBGeForce RTX 5090 D V2 24GBGeForce RTX 5090 D 32GBApple M5 Max 36GBApple M5 Pro 48GBApple M5 Max 64GBApple M3 Ultra 96GBApple M5 Max 128GBApple M2 Ultra 192GBApple M3 Ultra 256GBApple M3 Ultra 512GB

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 65.50 GiB. The real file is 69.75 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json
Layers
88
Attention heads
96
KV heads
8
Head dim
128
Hidden size
12288
Vocab
131,072
Sliding window
none
SWA period
—
MLA
no
Experts
—
Experts per token
—
use_sliding_window
—

Questions people ask

How much VRAM does Devstral-2-123B-Instruct-2512 need?
Q4_K_M is exactly 74,897,652,896 bytes (69.75 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Devstral-2-123B-Instruct-2512's KV cache?
11.00 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of Devstral-2-123B-Instruct-2512 should I use?
Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.