Computed

Dolphin Mistral 24B Venice Edition on a GeForce RTX 3060 12GB

Only technically. Q3_K_M weights fit in 11.0 GiB, but they leave room for just 723 tokens of context — below anything useful.

Change the numbers →Dolphin Mistral 24B Venice Edition on every card GeForce RTX 3060 12GB across every model

Verdict

Too tight

11.0 GiB usable of 12 GB

Weights (Q4_K_M)

13.5 GiB

44.7 GiB at FP16 · 24B params · 1 of 6 quant rows are measured files

KV cache

160.0 KiB per token at FP16

40 layers × 8 KV heads × 128 dimensions

Speed ceiling

30 tok/s

360 GB/s ÷ bytes read per token

Every quant, against this card

Weight bytes are the size of the real published file wherever one exists — 1 of these6 rows are measured from the published checkpoint, the rest computed from the parameter count. Max context is what the KV cache can grow to in whatever memory the weights leave behind, capped at the 128k tokens this model was trained to address.

Bars are the weight bytes at each precision; the dashed line is the usable memory of a GeForce RTX 3060 12GB. Drawn from the computed byte counts, not sketched.

Precision	Bits/weight	Weights	Source	Fits	Max context	Ceiling
FP16 / BF16	16.00	44.7 GiB	measured file	−33.7 GiB	—	2.2 tok/s offloaded
Q8_0	8.50	23.8 GiB	nominal	−12.7 GiB	—	5.0 tok/s offloaded
Q6_K	6.56	18.3 GiB	nominal	−7.3 GiB	—	7.4 tok/s offloaded
Q5_K_M	5.67	15.8 GiB	nominal	−4.8 GiB	—	9.5 tok/s offloaded
Q4_K_M	4.83	13.5 GiB	nominal	−2.5 GiB	—	13.0 tok/s offloaded
Q3_K_M	3.91	10.9 GiB	nominal	yes	723	30 tok/s

What the context actually costs

The KV cache is 2 · layers · kv_heads · head_dim bytes per token per element — 2 · 40 · 8 · 128 · 2 B = 160.0 KiB. It depends on the key/value head count, not the attention-head count and not the parameter count. This model shares 8 KV heads across 32 query heads, which divides the cache by 4 against multi-head attention.

Context	KV cache (FP16)	KV cache (8-bit)	Plus Q4_K_M weights
4,096	640 MiB	320 MiB	14.1 GiB
8,192	1.3 GiB	640 MiB	14.8 GiB
32,768	5.0 GiB	2.5 GiB	18.5 GiB
131,072	20.0 GiB	10.0 GiB	33.5 GiB

The speed ceiling, and where it comes from

Generating one token reads every active weight from memory once. At Q3_K_M that is 10.9 GiB, plus a pass over the KV cache. A GeForce RTX 3060 12GB moves 360 GB/s (192-bit × 15 Gbps), so the arithmetic ceiling is 30 tok/s. Treat it as a bound, not an estimate: attention overhead, kernel launches and imperfect memory access keep real runtimes at roughly 60–80% of it, and nothing pushes past it. On this pairing the ceiling is not the binding problem — the weights fit and the context does not, so the number above describes a model that can barely be given anything to read.

Where these numbers come from

The model

Parameters: 24,011,361,280
Layers: 40
Attention / KV heads: 32 / 8
Head dimension: 128
Trained context: 131,072
Checkpoint as published: 44.7 GiB

Read from dphn/Dolphin-Mistral-24B-Venice-Edition. The parameter count is the Hub's own total over the tensor shapes, not a figure taken from the model's name.

The accelerator

Memory: 12 GB GDDR6
Bandwidth: 360 GB/s
Bus: 192-bit × 15 Gbps
Assumed usable: 92% → 11.0 GiB

Capacity and bandwidth from the vendor's specification. The bandwidth figure is checked against the bus width and data rate it derives from. The usable fraction is an assumption, not a spec: a driver context, compute workspace and any attached display come out of the same pool before a weight is loaded.

Questions this pairing answers

How much VRAM does Dolphin Mistral 24B Venice Edition need?

44.7 GiB for the weights at FP16 — 24B parameters at two bytes each — and 13.5 GiB at Q4_K_M. The KV cache is on top of that and is not a fixed number: this model spends 160.0 KiB per token of context, so 8,192 tokens costs a further 1.3 GiB. A GeForce RTX 3060 12GB makes 11.0 GiB of its 12 GB available on the assumption below.

Can a GeForce RTX 3060 12GB run Dolphin Mistral 24B Venice Edition?

Only technically. Q3_K_M weights fit in 11.0 GiB, but they leave room for just 723 tokens of context — below anything useful. That is the weights and the KV cache together, against 11.0 GiB of usable memory.

How fast will Dolphin Mistral 24B Venice Edition run on a GeForce RTX 3060 12GB?

No faster than 30 tokens/second at Q3_K_M, and in practice below it. Decoding is memory-bound: every token reads all 10.9 GiB of weights plus the cache, and this card moves 360 GB/s. That division is the ceiling — no kernel, runtime or driver beats it, and a real runtime typically reaches 60–80% of it. That figure assumes the weights are resident, which here leaves almost no room for context — the speed is not the binding problem.

Why does the context length change how much memory Dolphin Mistral 24B Venice Edition needs?

Because the KV cache holds one key and one value vector per token, per layer, for the whole conversation, and it is allocated separately from the weights. This model has 40 layers and 8 key/value heads of 128 dimensions, shared across 32 query heads — grouped-query attention, which divides the cache by 4. That works out at 160.0 KiB per token. Parameter count tells you nothing about this number.

The same model on a different card

A different model on the same card

All 62 models on GeForce RTX 3060 12GB →

Method and limits. Weight bytes are the byte size of the real published file wherever one exists, and the model's exact parameter count times the published llama.cpp bits-per-weight where it does not. That distinction is on every row above and it matters at both ends: a sub-1B model's Q4_K_M file runs a third larger than the nominal figure because k-quants keep its embedding tables at higher precision, and an already-4-bit release cannot be quantized upward at all. The KV cache is 2 · layers · kv_heads · head_dim · bytes per token, summed over layers with each sliding-window layer capped at its window. The speed figure is a roofline bound, not a benchmark: bandwidth divided by bytes read per token, which no runtime exceeds and every runtime falls short of. The usable fraction of card memory is an assumption: 92% here, which is what this lane assumes for a dedicated card — the other class assumes 75%, so it is not one number applied to every device. That is the only assumed input on this page; every other figure is computed from the model config and the card's published specification. Nothing on this page is written by a language model. Architecture from the model's published config, fetched 2026-08-07. This checkpoint also carries a vision tower; its parameters are inside the totals above because they load whether or not you send it an image.