Reported by StorageReview SSD/NVMe/存储评测与行业新闻 on . Summarized and analyzed by XAIPART Editorial.
What happened
StorageReview published a hands-on review of Jan v0.8.6, an open-source desktop application marketed as a local ChatGPT alternative that runs downloaded GGUF models through llama.cpp and can also call hosted AI providers.S1
The test platform was an HP Z4 Rack G5 workstation running Windows 11 Pro for Workstations, equipped with an Intel Xeon w5-2465X (16 cores), 128GB of system RAM, and an NVIDIA RTX 5000 Ada Generation GPU with 32GB of VRAM, paired with NVIDIA driver 597.06 and CUDA 13.2 (nvidia-smi reading).S1 StorageReview installed Jan from scratch on that machine to evaluate setup flow, model management, GPU-accelerated inference, web search, multimodal input, and the local OpenAI-compatible API.S1
Jan v0.8.6 bundles llama.cpp directly inside the application (the review shows llama.cpp 0.5.0, build b11146), changes GPU-offload configuration, adjusts parallel-sequence defaults, and adds on-disk conversation prompt caching compared with earlier v0.8.5 and pre-0.8.5 builds.S1 StorageReview reran every performance test on v0.8.6 because of those changes.S1 The Windows installer required no account to use a local model, and on first launch Jan offered to download its recommended Jan V3.5 model, listed at 2.8GB.S1 The Hub also surfaced Jan V3.5 4B at 2.5GB in Q4_K_M and Jan V2 VL High at 4.7GB with multimodal and tools tags.S1
Using a RAID 5 vs RAID 6 prompt at an 8,192-token context, 100 GPU layers, and a batch size of 2,048, StorageReview measured a median decode rate of 123 tokens/sec for Jan V3.5 4B Q4_K_XL (503 output tokens, 4,839 MiB peak VRAM) and 82 tokens/sec for Jan V2 VL High Q4_K_M (1,728 output tokens, 8,197 MiB peak VRAM).S1 In a multimodal retest, Jan V2 VL High read six values from an nvidia-smi screenshot (GPU model, driver 597.06, 32,760 MiB VRAM, 1% utilization, 35°C, 4W of 250W) after 21 seconds of reasoning and 1,979 reported output tokens.S1 A built-in OpenAI-compatible server ran at http://127.0.0.1:1337/v1.S1
| Test GPU / VRAM | NVIDIA RTX 5000 Ada, 32GB VRAM | [S1] |
| System RAM | 128GB | [S1] |
| Jan V3.5 4B Q4_K_XL decode | 123 tok/s (median of 3 runs) | [S1] |
| Jan V3.5 4B peak VRAM | 4,839 MiB | [S1] |
| Jan V2 VL High Q4_K_M decode | 82 tok/s (median of 3 runs) | [S1] |
| Jan V2 VL High peak VRAM | 8,197 MiB | [S1] |
| Jan V3.5 4B context size used | 8,192 tokens; 100 GPU layers; batch size 2,048 | [S1] |
| Bundle of llama.cpp | v0.5.0, build b11146 | [S1] |
| Local API endpoint | http://127.0.0.1:1337/v1 | [S1] |
Background timeline
- - StorageReview publishes its Jan v0.8.6 hands-on on an HP Z4 Rack G5 with an RTX 5000 Ada, rerunning tests after v0.8.5 and v0.8.6 landed shortly before publication. [S1]
- - Jan v0.8.5 and v0.8.6 introduce bundled llama.cpp, revised GPU offload, adjusted parallel-sequence defaults, and on-disk conversation prompt caching. [S1]
XAIPART take
Technical impact
The published numbers show a small language model (4B parameters, Q4_K_XL) sitting comfortably inside 5GB of VRAM headroom on the RTX 5000 Ada, while a multimodal variant (Q4_K_M) used about 8GB at the same 8K context. That mapping is useful when sizing workstation GPUs for engineering teams that want on-device inference: a single 16GB RTX-class card can host the 4B model with room for context growth, while a 24GB card is the practical floor for the multimodal variant at long context. Bundling llama.cpp inside Jan also removes a class of DLL/CUDA runtime mismatches that procurement teams usually have to triage, which simplifies deployment images across a fleet. [XAIPART analysis]
The dialog around prompt caching is the second technical watchpoint: the review explicitly chose not to publish a prefill median because the cache was warm in repeated runs. For designers building internal tools, that means first-turn latency will be substantially higher than the steady-state decode numbers imply, and any internal benchmark should record cold vs. warm passes separately. [XAIPART analysis]
Market impact
The HP Z4 Rack G5 plus RTX 5000 Ada combination is positioned as a development workstation, and the StorageReview results act as a reference BOM for procurement: one Ada-generation 32GB GPU was enough to run both the text-only and multimodal variants tested. Because Jan is open source and the local API is OpenAI-compatible, design teams can prototype against http://127.0.0.1:1337/v1 and swap in hosted providers (OpenAI, Anthropic, Gemini, Groq, Mistral) without rewriting client code. For sourcing, the implication is that on-prem evaluation hardware can be justified at 32GB VRAM per seat for small-model local inference, rather than defaulting to cloud spend for every prompt. [XAIPART analysis]
The cited Jan V3.5 download at 2.8GB and Jan V2 VL High at 4.7GB indicate that even a single high-end workstation SSD can stage the model library; no specialty storage is required. That keeps the local-LLM workstation BOM close to a standard EDA or CAD seat, which matters for bulk purchasing decisions. [XAIPART analysis]
Action
Engineers evaluating local LLMs should treat the 123 tok/s and 82 tok/s figures as steady-state decode baselines on an RTX 5000 Ada and budget cold-context latency separately, since conversation prompt caching was enabled. Procurement should treat 32GB VRAM Ada-class GPUs as the practical reference point for teams planning to run both text-only and multimodal 4B-class models at 8K context. Before deployment, confirm that any external web-search provider (the Exa MCP service in the review hit its free rate limit) is on a paid tier or has documented headroom, and validate that the target client code works against Jan's OpenAI-compatible endpoint at port 1337 to avoid lock-in to a single vendor. [XAIPART analysis]
XAIPART Forecast
XAIPART expects that if StorageReview runs the same Jan V3.5 4B Q4_K_XL text-generation test on the RTX 5000 Ada within 6 months of this brief's publication (by 2027-04-14) and does not post a median decode rate at or above 123 tokens/sec at an 8,192-token context, then our forecast that the RTX 5000 Ada sets the Ada-generation local-LLM baseline will be wrong; if it clears 123 tok/s, the Ada 32GB card remains the procurement reference for 4B-class models at the same context size. The threshold (123 tok/s at 8,192 tokens) is taken directly from this brief's Key numbers, and the proper noun anchor is the NVIDIA RTX 5000 Ada Generation. [XAIPART analysis]
Facts are as reported by the cited source. The design-in, sourcing and forward-looking statements above are XAIPART's analysis, not the source's claims. Uncertainty: the 82 tok/s and 123 tok/s figures are single-source, single-hardware; Jan V2 VL High's output-token count was flagged by the source as including reasoning tokens and should not be compared directly with the 4B model; prefill performance was not published because of caching; and the Web Search feature could not complete in the retest because Exa's free MCP rate limit was hit.
Corrections: none at publication. This brief will be updated with a timestamp if the source corrects its report.
Alternatives and lifecycle
| Part | Relation | Lifecycle | Basis |
|---|---|---|---|
| — | No alternative on file NEEDS DATA | — | Requires match_type before any predecessor label may appear |
“Predecessor” and “alternative” labels come only from a recorded match type, a lifecycle status, or the source text. Where no match is recorded, no label is shown. Our own listings are not a market reference.
Affected parts
No catalog match. This story does not map to any part currently listed on XAIPART. We do not infer matches, and we do not show a placeholder list.
Cost pressure by dimension
| Dimension | FRED series ID | Unit | Lag | As of |
|---|---|---|---|---|
| Input cost | PCOPPUSDM, PALUMUSDM, PNICKUSDM | USD/mt | 96 d | |
| Supply/demand | PCU334413334413, IPG3344S | index(1982=100) / index(2017=100) | 46 d | |
| FX & energy | DEXCHUS, DCOILBRENTEU, DHHNGSP | rate / USD/bbl / USD/MMBtu | 6 d |
No free authoritative interface for: gold · silver · rare earth · polysilicon · lithium carbonate · market share — listed rather than approximated.
How this was produced
Process evidence — fetch, source sanity, model and gate
| Source fetched | StorageReview SSD/NVMe/存储评测与行业新闻 · www.storagereview.com · 27,227 chars · fetch status success |
|---|---|
| Source sanity (S1) | ok — passed before any model call (no s1_ block recorded) |
| Model / prompt | MiniMax-M3 · prompt Blog Agent - News Summary (v2.8-final) · 14,530 tokens · 41.2s · run completed |
| Gate | verdict pass · 6/6 applicable checks passed · spec v2.9 · C10 ratio 0.00425 · C10 rounds 0 |
| Time contract | 3 <time datetime> values in this brief — 0 violations (+08:00 or date-only) |
Every row above is produced by the pipeline, not written by hand.
Comments
Be the first to comment.
🔐 Please sign in to post a comment.
Don't have an account? Sign up