What GPU does StarCoder2 15B need?

A 12GB card. Measured peak was ~11GB at Q4_K_M. That undercuts Codestral 22B's ~14GB floor by a full card class: B580/RTX 3060 12GB territory instead of 16GB.

Why does its training data matter?

BigCode documents it fully: permissively-licensed code from The Stack v2, license-filtered, developer opt-outs honored. For organizations with IP concerns about AI-generated code, that provenance is the model's headline feature.

StarCoder2 15B or Codestral 22B?

Codestral completes better in our experience. It's the completion specialist pick if 16GB is available. StarCoder2 counters with a lighter floor (12GB cards) and auditable training data. Both are FIM-native; the tradeoff is capability versus footprint-plus-provenance.

Is it fast enough for editor completions?

Yes: completion UX depends on prompt ingestion plus first tokens, and it measured up to 4,428 prompt t/s with 144 tok/s generation. Even mid-tier cloud cards keep suggestions instant-feeling.

What quantization did you test?

Q4_K_M via llama.cpp, the same pinned harness as every LLM here: so its numbers compare directly against Codestral, the Qwen coders and everything else in the database.

How do I pair it in a local coding stack?

Use it as the completion engine next to a bigger agent model: StarCoder2 answers the editor's fill-in-middle requests from a 12GB card while something like Devstral 24B or Qwen3 Coder 30B-A3B handles chat, refactors and agent loops. The split mirrors how the models were trained, and a 24GB card hosts both halves.

StarCoder2 15B · 11 GPUs measured first-party · llama.cpp Q4_K_M · Updated July 2026

What GPU Do You Need for StarCoder2 15B?

StarCoder2 15B is the BigCode project's open code model, trained on permissively-licensed, documented source data with fill-in-the-middle as a first-class objective. We measured it on 11 GPUs (llama.cpp, Q4_K_M): 144 tok/s on the B300, ~11GB peak VRAM.

Benchmarked weights: bartowski/starcoder2-15b-instruct-v0.1-GGUF

144.05tok/s

Fastest: NVIDIA B300

measured, 3-run llama-bench

~11GB

VRAM needed (measured peak)

GPU-independent, applies to every card

GPUs measured

same pinned harness

0.74tok/W

Most efficient: NVIDIA H100 80GB HBM3

real power sampling, not TDP

What GPU Do You Need for StarCoder2 15B?, tok/s, fastest 11

NVIDIA B300

144.05 tok/s

NVIDIA B200

138.64 tok/s

NVIDIA H200

136.29 tok/s

NVIDIA H100 80GB HBM3

134.92 tok/s

NVIDIA RTX PRO 6000 Blackwell Workstation Edition

122.09 tok/s

NVIDIA A100 40GB SXM4

80.6 tok/s

NVIDIA A100 80GB SXM4

79.22 tok/s

NVIDIA L40S

66.85 tok/s

NVIDIA A10G

42.92 tok/s

NVIDIA L4

24.39 tok/s

NVIDIA T4

14.42 tok/s

Measured on our own bench. A card absent from this chart has not been run on this model yet, or cannot fit it.

StarCoder2 15B. Measured generation speed by GPU

NVIDIA B300144.05

NVIDIA B200138.64

NVIDIA H200136.29

NVIDIA H100 80GB HBM3134.92

NVIDIA RTX PRO 6000 Blackwell Workstation Edition122.09

NVIDIA A100 40GB SXM480.6

NVIDIA A100 80GB SXM479.22

NVIDIA L40S66.85

NVIDIA A10G42.92

NVIDIA L424.39

NVIDIA T414.42

GPU	tok/s	Prompt t/s	tok/W	Avg power
NVIDIA B300	144.05	2456.5	0.44	329.9 W
NVIDIA B200	138.64	4428.4	0.4	344.6 W
NVIDIA H200	136.29	4116.4	0.64	213.5 W
NVIDIA H100 80GB HBM3	134.92	4072.8	0.74	181.3 W
NVIDIA RTX PRO 6000 Blackwell Workstation Edition	122.09	6106.1	0.58	210.8 W
NVIDIA A100 40GB SXM4	80.6	2189.5	0.56	143.0 W
NVIDIA A100 80GB SXM4	79.22	2261	0.44	178.1 W
NVIDIA L40S	66.85	4647.1	0.41	161.1 W
NVIDIA A10G	42.92	1524.1	0.41	105.9 W
NVIDIA L4	24.39	1271.2	0.44	55.9 W
NVIDIA T4	14.42	548	0.24	59.3 W

The open-data coder. StarCoder2's distinction mirrors SmolLM3's: the BigCode initiative documents exactly what code it trained on (The Stack v2, license-filtered, with opt-outs honored). For teams that care where their completion model's training code came from, legal departments do, it's the most defensible coder in our lineup. Technically it's a fill-in-the-middle native at a size between our other code picks: bigger than Qwen2.5-Coder 7B, lighter than Codestral 22B.

Where it lands on hardware. The ~11GB floor makes 12GB cards its natural home, an Arc B580 or RTX 3060 12GB hosts it where Codestral's ~14GB wants 16GB. Speed is honest for the size: 144 tok/s peak, 136 on the H200, and a usable 24 tok/s even on an L4. For completions, remember prompt ingestion dominates perceived latency: 2,457-4,428 prompt t/s at the top of the chart means large-file context loads fast.

Our verdict

StarCoder2 15B: 144 tok/s peak, ~11GB floor, the documented-training-data coder that fits a 12GB card. Between Qwen's 7B and Codestral's 22B on size, and ahead of both on provenance; if your completion model needs a paper trail, this is it.

FAQ

What GPU does StarCoder2 15B need?: A 12GB card. Measured peak was ~11GB at Q4_K_M. That undercuts Codestral 22B's ~14GB floor by a full card class: B580/RTX 3060 12GB territory instead of 16GB.
Why does its training data matter?: BigCode documents it fully: permissively-licensed code from The Stack v2, license-filtered, developer opt-outs honored. For organizations with IP concerns about AI-generated code, that provenance is the model's headline feature.
StarCoder2 15B or Codestral 22B?: Codestral completes better in our experience. It's the completion specialist pick if 16GB is available. StarCoder2 counters with a lighter floor (12GB cards) and auditable training data. Both are FIM-native; the tradeoff is capability versus footprint-plus-provenance.
Is it fast enough for editor completions?: Yes: completion UX depends on prompt ingestion plus first tokens, and it measured up to 4,428 prompt t/s with 144 tok/s generation. Even mid-tier cloud cards keep suggestions instant-feeling.
What quantization did you test?: Q4_K_M via llama.cpp, the same pinned harness as every LLM here: so its numbers compare directly against Codestral, the Qwen coders and everything else in the database.
How do I pair it in a local coding stack?: Use it as the completion engine next to a bigger agent model: StarCoder2 answers the editor's fill-in-middle requests from a 12GB card while something like Devstral 24B or Qwen3 Coder 30B-A3B handles chat, refactors and agent loops. The split mirrors how the models were trained, and a 24GB card hosts both halves.