Released 2026-09-10 · general agent
Deployment evidence
DeepSeek · 552B backbone + 196B Engram memory total · 8B per prefill token / 16B per decode token active · 1M tokens
Best-fit scenarios
Frontier-scale private multimodal coding, document and long-context agent workloads where a 500-plus-GB checkpoint and dedicated serving build are acceptable.
Not recommended
- Treating the 8B/16B active-parameter figures as weight-memory requirements; the full checkpoint, Engram tables and DSpark drafter still have to be loaded
- Assuming a successful text-only GB200 recipe also validates the separate image encoder and multimodal request path
- Launching a public API without pinning the dedicated prompt encoder, tool parser and reasoning-effort mapping
- Sizing full-context concurrency from the one-million-token specification without workload measurements
Long-context coding and tool-using agents
Vendor-statedDeepSeek positions V4.1 Flash for coding, reasoning and agent tasks with a one-million-token context and continuous reasoning-effort control. The published quality results are vendor evaluations, not ChinaAPI reproductions.
Capability sourceImage-grounded coding and document analysis
Vendor-statedThe official checkpoint includes a native vision encoder and accepts image-plus-text prompts, making it a candidate for screenshot, UI and document workflows.
Capability sourcePrivate multimodal agent endpoint
ChinaAPI inferenceThe open weights and documented OpenAI-compatible serving path make an operator-controlled endpoint plausible when data locality justifies data-center-class hardware; this is a deployment fit assessment, not a capacity result.
Official launch / minimum
Evidence BOfficial capacity floor
vLLM estimates about 476 GiB of checkpoint data and a 614 GB planning minimum including 20% headroom. Its current recipe fits the model on one 4x GB200 NVL4 tray at TP4 or one 8x H200 node. This is a framework capacity floor, not a throughput or concurrency result. Primary recipe
200-person team
Evidence BValidated worker benchmark required
For the defined 200-seat profile, start with one measured 4x GB200 or 8x H200 worker and replay the real text/image, context and reasoning-effort mix. Add an independent worker when maintenance or failure recovery is required. No official recipe maps employee count to capacity. Primary recipe
Commercial API
Evidence BOfficial text recipe multimodal validation required
vLLM publishes a verified text-only TP4 GB200 worker and a 1P1D layout that uses one 4-GPU GB200 tray per role. A commercial service still needs independent redundant workers, admission control, pinned encoding/parsers, and separate image-path smoke and load tests. Primary recipe
Commercial-use check
MIT
MaaS / hosted service: No model-specific MaaS restriction identified in the MIT license published with the official weights.
Attribution: Retain the copyright and permission notice in copies or substantial portions of the weights. The weights are provided without warranty.
Known limitations and open questions
- No ChinaAPI hardware reproduction, quality benchmark or like-for-like comparison between the downloadable checkpoint and hosted route
- The current vLLM path requires the dedicated vLLM 0.30 day-zero Docker image; no pip wheel serves this architecture at the evidence cutoff
- The verified GB200 TP4 and 1P1D runs are text-only and do not by themselves validate image input
- The release uses a dedicated prompt encoder rather than a Jinja chat template, and open-weight reasoning-effort labels do not map identically to the hosted DeepSeek API
- DSpark acceptance, long first-load time and full-context concurrency remain workload- and infrastructure-dependent