Aventail™ Link vs Ollama
- Model
- Gemma 4 E2B-it
- Quantisation
- Q4_K_M
- Load
- Single stream
- Measured
Both runtimes served the same Gemma 4 E2B-it Q4_K_M model, one request at a time, with identical sampling. These results were measured and originally published in June 2026 under the product’s former name, Locai Link. They have not been rerun on current releases and describe only the machines and settings recorded here, not performance in general.
Showing Linux conversation results.
Linux · NVIDIA RTX 4070Conversation
- Prompt
- ~30 tokens
- Reply
- ~300 tokens
- Runs
- 20 (+3 warmup)
- Errors
- 0
Short conversational prompt (~30 tokens), mid-length reply (~300 tokens).
- Aventail™ Link measured better
- time to first token, median; time to first token, p95; prefill throughput; decode throughput, median; decode throughput, p95; end-to-end latency, median
- Ollama measured better
- peak VRAM; peak host RSS
Time to first token, median (ms) lower is better
- Link
- 31 ms
- Ollama
- 326 ms
Decode throughput, median (tok/s) higher is better
- Link
- 196 tok/s
- Ollama
- 166 tok/s
End-to-end latency, median (s) lower is better
- Link
- 1.6 s
- Ollama
- 2.1 s
Peak host RSS (MB) lower is better
- Link
- 5,799 MB
- Ollama
- 3,022 MB
| Metric | Aventail™ Link | Ollama | Link − Ollama |
|---|---|---|---|
| Time to first token, median (ms)lower is better | 31 | 326 | −295 ms (Link better) |
| Time to first token, p95 (ms)lower is better | 132 | 338 | −206 ms (Link better) |
| Prefill throughput (tok/s)higher is better | 936 | 92 | +844 tok/s (Link better) |
| Decode throughput, median (tok/s)higher is better | 196 | 166 | +30 tok/s (Link better) |
| Decode throughput, p95 (tok/s)higher is better | 198 | 169 | +29 tok/s (Link better) |
| End-to-end latency, median (s)lower is better | 1.6 | 2.1 | −0.5 s (Link better) |
| Peak VRAM (MiB)lower is better | 4,244 | 1,990 | +2,254 MiB (Ollama better) |
| Peak host RSS (MB)lower is better | 5,799 | 3,022 | +2,777 MB (Ollama better) |
Linux · NVIDIA RTX 4070Document analysis
- Prompt
- ~9,500 tokens
- Reply
- ≤500 tokens
- Runs
- 10 (+3 warmup)
- Errors
- 0
Long-context document prompt (~9,500 tokens) from 'From the Earth to the Moon', mid-length reply (≤500 tokens).
- Aventail™ Link measured better
- time to first token, median; prefill throughput; decode throughput, median; decode throughput, p95; end-to-end latency, median
- Ollama measured better
- peak VRAM; peak host RSS
Time to first token, median (ms) lower is better
- Link
- 56 ms
- Ollama
- 385 ms
Decode throughput, median (tok/s) higher is better
- Link
- 174 tok/s
- Ollama
- 158 tok/s
End-to-end latency, median (s) lower is better
- Link
- 1.2 s
- Ollama
- 1.6 s
Peak host RSS (MB) lower is better
- Link
- 6,003 MB
- Ollama
- 3,091 MB
| Metric | Aventail™ Link | Ollama | Link − Ollama |
|---|---|---|---|
| Time to first token, median (ms)lower is better | 56 | 385 | −329 ms (Link better) |
| Time to first token, p95 (ms)lower is better | Not reported | Not reported | Not reported |
| Prefill throughput (tok/s)higher is better | 169,466 | 24,728 | +144,738 tok/s (Link better) |
| Decode throughput, median (tok/s)higher is better | 174 | 158 | +16 tok/s (Link better) |
| Decode throughput, p95 (tok/s)higher is better | 178 | 160 | +18 tok/s (Link better) |
| End-to-end latency, median (s)lower is better | 1.2 | 1.6 | −0.4 s (Link better) |
| Peak VRAM (MiB)lower is better | 4,266 | 2,006 | +2,260 MiB (Ollama better) |
| Peak host RSS (MB)lower is better | 6,003 | 3,091 | +2,912 MB (Ollama better) |
Memory trade-off: Ollama used less peak VRAM and host memory than Link in both scenarios.
Both workloads on Linux compared
- Time to first token, median, Link vs Ollama: 31 vs 326 ms (conversation); 56 vs 385 ms (document analysis). Link better in both scenarios; lower is better.
- Decode throughput, median, Link vs Ollama: 196 vs 166 tok/s (conversation); 174 vs 158 tok/s (document analysis). Link better in both scenarios; higher is better.
- End-to-end latency, median, Link vs Ollama: 1.6 vs 2.1 s (conversation); 1.2 vs 1.6 s (document analysis). Link better in both scenarios; lower is better.
- Peak VRAM, Link vs Ollama: 4,244 vs 1,990 MiB (conversation); 4,266 vs 2,006 MiB (document analysis). Ollama better in both scenarios; lower is better.
- Peak host RSS, Link vs Ollama: 5,799 vs 3,022 MB (conversation); 6,003 vs 3,091 MB (document analysis). Ollama better in both scenarios; lower is better.
The original report attributes the narrower decode gap to both backends sharing the same underlying CUDA kernels for the generation loop.
On memory, the original web page says Ollama defers KV-cache and weight allocations more aggressively while Locai Link pre-allocated them; the PDF report describes the gap as different KV-cache strategies, with Link allocating a larger fixed cache.
Linux environment and source notes (includes a build discrepancy)
Source discrepancy: llama.cpp build: the original web page data records b9789, while the original raw JSON and the PDF report both record b9222. We have not chosen between them, so the exact Link build used for these results cannot be confirmed.
Machine
- Operating system
- Arch Linux
- GPU
- NVIDIA RTX 4070, 12,288 MiB VRAM
- GPU driver
- 610.43.02
- CUDA
- 13.3
- Compiler
- nvcc / g++-15
- Reasoning
- Off (LLAMA_ARG_REASONING=off / think: false)
Aventail™ Link (then Locai Link)
- Backend
- llama.cpp, built from source (build number conflicts between sources: b9789 or b9222)
- Compute
- CUDA
- API
- /v1/chat/completions, OpenAI-compatible SSE
Ollama
- Package
- ollama-cuda (Arch package)
- Settings
- OLLAMA_LLM_LIBRARY=cuda_v13, num_gpu=999 (request GPU offload)
- Compute
- CUDA
- API
- /api/chat, NDJSON
Not specified in the source: Ollama release version.
Original note: Single-stream measurements only. One request in flight at a time per backend. Concurrent load was not measured. Results are specific to Arch Linux and the ollama-cuda Arch package; build steps and library paths differ on Ubuntu and RHEL.
macOS · Apple M2, 16 GBConversation
- Prompt
- ~30 tokens
- Reply
- ~300 tokens
- Runs
- 20 (+3 warmup)
- Errors
- 0
Short conversational prompt (~30 tokens), mid-length reply (~300 tokens).
- Aventail™ Link measured better
- time to first token, median; time to first token, p95; prefill throughput
- Ollama measured better
- decode throughput, median; decode throughput, p95; end-to-end latency, median; peak host RSS
Time to first token, median (ms) lower is better
- Link
- 205 ms
- Ollama
- 478 ms
Decode throughput, median (tok/s) higher is better
- Link
- 39 tok/s
- Ollama
- 41 tok/s
End-to-end latency, median (s) lower is better
- Link
- 7.9 s
- Ollama
- 7.8 s
Peak host RSS (MB) lower is better
- Link
- 6,493 MB
- Ollama
- 3,332 MB
| Metric | Aventail™ Link | Ollama | Link − Ollama |
|---|---|---|---|
| Time to first token, median (ms)lower is better | 205 | 478 | −273 ms (Link better) |
| Time to first token, p95 (ms)lower is better | 215 | 490 | −275 ms (Link better) |
| Prefill throughput (tok/s)higher is better | 141 | 60 | +81 tok/s (Link better) |
| Decode throughput, median (tok/s)higher is better | 39 | 41 | −2 tok/s (Ollama better) |
| Decode throughput, p95 (tok/s)higher is better | 40 | 41 | −1 tok/s (Ollama better) |
| End-to-end latency, median (s)lower is better | 7.9 | 7.8 | +0.1 s (Ollama better) |
| Peak VRAM (MiB)lower is better | Not applicable — unified memory | Not applicable — unified memory | Not applicable |
| Peak host RSS (MB)lower is better | 6,493 | 3,332 | +3,161 MB (Ollama better) |
macOS · Apple M2, 16 GBDocument analysis
- Prompt
- ~9,500 tokens
- Reply
- ≤500 tokens
- Runs
- 10 (+3 warmup)
- Errors
- 0
Long-context document prompt (~9,500 tokens) from 'From the Earth to the Moon', mid-length reply (≤500 tokens).
- Aventail™ Link measured better
- time to first token, median; prefill throughput; decode throughput, p95
- Ollama measured better
- end-to-end latency, median; peak host RSS
- Equal
- decode throughput, median
Time to first token, median (ms) lower is better
- Link
- 332 ms
- Ollama
- 589 ms
Decode throughput, median (tok/s) higher is better
- Link
- 30.4 tok/s
- Ollama
- 30.4 tok/s
End-to-end latency, median (s) lower is better
- Link
- 5.5 s
- Ollama
- 5.3 s
Peak host RSS (MB) lower is better
- Link
- 6,759 MB
- Ollama
- 3,429 MB
| Metric | Aventail™ Link | Ollama | Link − Ollama |
|---|---|---|---|
| Time to first token, median (ms)lower is better | 332 | 589 | −257 ms (Link better) |
| Time to first token, p95 (ms)lower is better | Not reported | Not reported | Not reported |
| Prefill throughput (tok/s)higher is better | 28,693 | 16,178 | +12,515 tok/s (Link better) |
| Decode throughput, median (tok/s)higher is better | 30.4 | 30.4 | No difference |
| Decode throughput, p95 (tok/s)higher is better | 32.3 | 31.9 | +0.4 tok/s (Link better) |
| End-to-end latency, median (s)lower is better | 5.5 | 5.3 | +0.2 s (Ollama better) |
| Peak VRAM (MiB)lower is better | Not applicable — unified memory | Not applicable — unified memory | Not applicable |
| Peak host RSS (MB)lower is better | 6,759 | 3,429 | +3,330 MB (Ollama better) |
Memory trade-off: Ollama’s peak memory was roughly half of Link’s in both scenarios. On Apple Silicon the GPU shares system memory, so there is no separate VRAM figure; host RSS captures the full footprint.
Both workloads on macOS compared
- Time to first token, median, Link vs Ollama: 205 vs 478 ms (conversation); 332 vs 589 ms (document analysis). Link better in both scenarios; lower is better.
- Decode throughput, median, Link vs Ollama: 39 vs 41 tok/s (conversation); 30.4 vs 30.4 tok/s (document analysis). conversation: Ollama better; document analysis: equal; higher is better.
- End-to-end latency, median, Link vs Ollama: 7.9 vs 7.8 s (conversation); 5.5 vs 5.3 s (document analysis). Ollama better in both scenarios; lower is better.
- Peak host RSS, Link vs Ollama: 6,493 vs 3,332 MB (conversation); 6,759 vs 3,429 MB (document analysis). Ollama better in both scenarios; lower is better.
macOS environment and source notes
Machine
- Operating system
- macOS
- Hardware
- Apple M2 SoC · 16 GB unified memory
- Compute
- Metal (Apple Silicon)
- Reasoning
- Off (LLAMA_ARG_REASONING=off / think:false)
Aventail™ Link (then Locai Link)
- Backend
- llama.cpp (Metal)
- Compute
- Metal
- API
- /v1/chat/completions, OpenAI-compatible SSE
Ollama
- Package
- ollama (native macOS build)
- Compute
- Metal
- API
- /api/chat, NDJSON
Not specified in the source: Ollama release version, macOS version, Locai Link / llama.cpp build (the report says only that Locai Link’s installer downloaded a Metal-accelerated llama.cpp prebuilt).
Original note: Single-stream measurements on Apple Silicon (M2, 16 GB unified memory). Concurrent load not measured. On Apple Silicon the GPU shares the system memory pool — there is no separate VRAM, so RSS captures the full memory footprint.
Windows
Windows: no published results in this dataset. The Linux results should not be read as a prediction of Windows performance.
Method and downloads
Original files
Unmodified copies of the files published with the June 2026 results. The macOS results were published as a report only; no macOS raw run data is available.
- Original Linux report (PDF)locai-link-vs-ollama-report.pdf · 490 KB
- Original Linux raw data (JSON)ollama-vs-locai-cuda-linux-raw-data.json · 4 KB
- Original macOS report (PDF)locai-link-vs-ollama-mac-report.pdf · 288 KB
Shared settings and limits
- Model
- Gemma 4 E2B-it, 4.6B parameters, Q4_K_M
- File
- GGUF, 2.9 GB
- Architecture
- Sliding-window attention, shared KV layers
- Sampling
- temperature = 0, top_p = 1, seed = 42, streaming = true
- Load
- Single stream: one request in flight at a time per backend. Concurrent load was not measured.
Output equivalence
The original report records: “Token-identical replies at temperature=0 confirmed for all prompts across both backends.” This applies to the tested prompts only and is not a general guarantee of output quality.
Source checks and discrepancies
Every value on this page matches the original web page data. The Linux raw JSON and the Linux and macOS PDF reports carry the same values. The reports’ ratios are derived from these values, and their distribution and per-run charts are images without published figures, so neither adds new measurements here. No benchmarks were rerun for this page.
- Linux: llama.cpp build: the original web page data records b9789, while the original raw JSON and the PDF report both record b9222. We have not chosen between them, so the exact Link build used for these results cannot be confirmed.
- The original web page omits the reasoning setting for Linux; the raw JSON and report record it as off on both backends.
- The measurement date is June 2026, as recorded in the sources. The upload date of the macOS report file is not used as a measurement date.