NVIDIA GeForce RTX 3070 Ti | VRAM Check Result

Leaderboard result

419.1tok/sCanonical score

ClassS

Public class

Placed #5 / 37 globally under the canonical benchmark profile.

Best fit today: Coding · General chat · Math / reasoning · Embeddings.

NVIDIA GeForce RTX 3070 Ti

ModelTinyLlama 1.1B (Q4_K_M)

QuantQ4_K_M

DateMar 12, 2026

Global standingPublic leaderboard position

Global#5/ 37 Cohort#1/ 1 Country#1/ 1

Workload & profileMeasured benchmark configuration

ModelTinyLlama 1.1B (Q4_K_M)

MethodQ4_K_M

ProfileCanonical profile

Recommended workloadsBest fit today

Coding

General chat

Math / reasoning

Embeddings

Use the sections below if you want the full fit matrix and planner-backed next step.

Details & explanationContext and scoring rules

Run meaningOfficial scored run

This run sets an official score, class, and public ranking under the canonical benchmark profile.

Trust & system contextVerified GPU (Leaderboard Eligible)

This is the canonical number the leaderboard uses for this run.

vulkan backendDiscrete GPU memoryFirst limit: Memory headroom

Step 01

Measured output

The official score plus the few supporting numbers that explain how this run behaved.

Decode ThroughputScored pass

419.1tok/s

This is the canonical number the leaderboard uses for this run.

ProfileTinyLlama 1.1B (Q4_K_M)

Standing#5 / 37 global

Prefill throughput

19766.2 tok/s

TTFT p50

34 ms

Effective model memory

6.3 GB

Scored pass duration

Stability rangeDecode / TTFT

1.03% / 0.78%

Step 02

What this run supports

This is the plain-language support map for the current build: what runs well now, what is close, and what still needs bigger hardware.

Ready now4 workload lanes run well

Coding · General chat · Math / reasoning · Embeddings are the clearest comfortable fits on this machine today.

CodingGeneral chatMath / reasoningEmbeddings

Closest stretchNothing is one small step away

The current build already covers the best near-term lanes, so there is no obvious single stretch target.

No clear stretch lane

Still outside envelopeNo lane is fully off-limits

Nothing here is completely out of reach at the workload-lane level.

No hard no lane

Using the local model catalog for this guidance.

CodingRuns well now

25% of tracked models fit now

Works well nowCoding works well on this build

Good starting points: Granite Code 3B · StarCoder2 3B · Qwen3 4B Instruct.

What more headroom changesAn upgrade adds breathing room

More headroom would make this lane feel safer, but it would not unlock a new class of models yet.

Main limit under reviewBased on local model catalog

General chatRuns well now

5% of tracked models fit now

Works well nowGeneral chat works well on this build

Good starting points: Llama 3.2 1B Instruct · TinyLlama 1.1B · Falcon3 1B Instruct.

What more headroom changesAn upgrade adds breathing room

More headroom would make this lane feel safer, but it would not unlock a new class of models yet.

Main limit under reviewBased on local model catalog

Math / reasoningRuns well now

4% of tracked models fit now

Works well nowMath / reasoning works well on this build

Good starting points: Llama 3.2 1B Instruct · TinyLlama 1.1B · Falcon3 1B Instruct.

What more headroom changesAn upgrade adds breathing room

More headroom would make this lane feel safer, but it would not unlock a new class of models yet.

Main limit under reviewBased on local model catalog

EmbeddingsRuns well now

100% of tracked models fit now

Works well nowEmbeddings works well on this build

Good starting points: all-minilm · Nomic Embed Text v1.5 · mxbai-embed-large.

What more headroom changesAn upgrade adds breathing room

More headroom would make this lane feel safer, but it would not unlock a new class of models yet.

Main limit under reviewBased on local model catalog

Step 03

Primary constraint

This is the clearest limit in the current build, how we know it, and why the recommendation points there.

Primary constraintMemory is the first thing this build runs out of

This run had 6.3 GB of effective model memory. That is enough for smaller local AI workloads, but larger coding and reasoning models run out of room before raw speed becomes the problem.

What hits first6.3 GB effective model memory

Bigger coding and reasoning families run out of room before speed becomes the main issue.

What still worksCoding · General chat · Math / reasoning · Embeddings

Smaller local AI workloads are already comfortable on this build today.

Why we recommend thisPrioritize more memory headroom

That is the fastest way to move from smaller models to larger ones without changing the whole character of the build.

Why this recommendationMore memory headroom unlocks the next tier of models

That is why the recommendation points to more memory headroom first. It unlocks larger model families sooner than chasing speed alone.

No workload lane is fully blockedThis is why Step 04 centers the NVIDIA ladder on Keep current GPU.

Trust and system context

Compact evidence behind why this run should be believed and how the machine was identified.

Execution profileVerified GPU benchmark

Validation statusVerified

Run qualityStable

Methodologyv1.1

GPUNVIDIA GeForce RTX 3070 Ti

CPU12th Gen Intel(R) Core(TM) i7-12700F

VRAM / RAM7.8 / 63.8 GB

OSWindows 11 24H2 (Build 26200)

Client version0.1.0

DriverNot reported

Step 04

GPU upgrade stage

A build-relative ladder that stays practical first, then shows the sweet spot, then the aspirational ceiling.

Detected bottleneckMemory is the first thing this build runs out of

That is why the recommendation points to more memory headroom first. It unlocks larger model families sooner than chasing speed alone.

Minimum acceptableNo smaller step clearly stands out

Hardware path

Current buildHold

Stay with the current build

Why this tierBudget-conscious jump

Choose this if you want the lowest-cost GPU jump that still changes what the build can realistically run.

What it opensUse this result as the baseline

The measured envelope does not show a modest upgrade that clearly changes the result enough to recommend it by default.

Open GPU catalog

RecommendedKeep current GPU

Hardware pathDefault choice

Current buildHold

Stay with the current build

Why this tierBest long-term value

Choose this if you want the strongest day-to-day value. It is the most balanced step for real gains without jumping straight into premium pricing.

What it opensOpen the planner if you want to compare targeted paths

There is no single obvious purchase that clearly earns a default recommendation from this run.

Open GPU catalog

Top endNo larger normal step stands out

Hardware path

Current buildHold

Hold the current build

Why this tierPeak consumer path

Choose this only if budget is flexible and you want the highest normal consumer ceiling, not the best value per dollar.

What it opensOnly move higher if you have a very specific target

At the top of the normal consumer path, the benchmark still does not show a clearly better next purchase.

Open GPU catalog

Keep this result movingCompare this build on the leaderboard or share the result without losing context.

See leaderboard context Share result

Optional depth

Planner, catalog, and audit trail

The report above stays focused on the outcome. Open this workspace when you want to compare model families, browse the tracked catalog, or inspect the pass evidence behind the score.

Open this only when you need the deeper planning layer

25% of tracked coding fits clear on this machine today.