MoE workstation671B total, 37B active per token. Q4_K_M requires ~360GB RAM/VRAM. Use Q2_K (~180GB) for DGX Spark.
- 671.0B
- 128k
- mit
- Mar 2025
Scores #
Score per dollar #
363 pts per $/M input
general_score (87) divided by cheapest input price
($0.24/M).
Higher is better value. [See live pricing](/models/deepseek-v3-0324/pricing).
Run it locally #
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
Can you run it? - reference rigs
| Rig | Q2_K | Q4_K_M |
|---|---|---|
| NVIDIA Jetson Orin NX 16GB | ||
no -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudno -> cloudFit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
How can a 24GB GPU run a 744B model? It does not load the model into VRAM. The quantized weights (e.g. ~410GB at Q4) sit in system RAM; the GPU only holds the small shared attention and router tensors and accelerates prompt processing. Because GLM 5.2 is a Mixture-of-Experts model, each token activates only ~40B of its 744B params, so llama.cpp streams just those active experts from system RAM to the GPU each token (the -cmoe
offload path).
That makes decode speed bound by system-RAM bandwidth, not GPU bandwidth - single digits on DDR4, which is why these rigs show 3-8 t/s even though they “fit.” A bigger GPU (e.g. 2x 3090) keeps more experts resident on-card and raises tok/s; a smaller GPU still runs it but pays the bandwidth tax. A 744B dense model could not run this way - only MoE’s small-active-params trick makes it possible.
Aggressive quants (1-2 bit) trade accuracy for size - roughly 17% accuracy loss at 2-bit vs full precision, and real long-context work often needs Q5 or Q6 even when lower quants “fit.”
Formula estimates here are conservative; real tuned setups can exceed them (one HN user reports ~6 tok/s on a 512GB DDR4 + 2x 3090 rig).
Download options #
Or run it in the cloud #
Live per-provider pricing, throughput and uptime - refreshed about 1 hour ago via OpenRouter. Click a column to sort.
| Provider | Type | Input $/M | Output $/M | Cache $/M | Tok/s | Latency | Uptime | Value |
|---|---|---|---|---|---|---|---|---|
| SiliconFlow | ||||||||
| API | 0.25 | 1.00 | 0.135 | - | - | 98.89% | best uptime | |
| Novita | ||||||||
| API | 0.27 | 1.12 | 0.135 | - | - | 97.79% | ||
| DeepInfra | ||||||||
| avoid | ||||||||
| API | 0.24 | 0.90 | 0.135 | - | - | 52.25% | cheapest |
Default order: throughput among 95%+ uptime providers, then latency; subscriptions last. Sort by any column. Subscription rows show $/mo in the Value column - per-token columns are "-". Affiliate links are marked sponsored / nofollow. Confirm current pricing on the provider's site before committing.
[Detailed API pricing page + JSON endpoint →](/models/deepseek-v3-0324/pricing)
[See who runs DeepSeek in production →](/adoption/deepseek)
Inference cost over time #
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.