DeepInfer Sources: Official Docs and V4.1-Flash References

Every factual claim in this dossier about DeepInfer, DeepSeek's V4.1-Flash inference framework, is tied to a public source. We separate four layers: A official primary sources, B official claims that cannot be independently verified, C named secondary coverage, and D community observations that change by the hour. A number is only as strong as its layer.

A. Official primary sources

• DeepInfer repository — README, license, feature list, hardware, build requirements, performance charts. Accessed 2026-10-20.
• DeepInfer deployment Q&A — production architecture, Wide-EP, EPLB, KV cache, DSpark, scheduling, Recipe. Updated 2026-10-20 11:03 UTC.
• Examples README — offline inference, benchmark, sweep parameters, CP/EP flags.
• Eval README — AIME 2026, GPQA Diamond, HLE reference scores and protocol.
• deepseek-recipe — protocol conversion, tokenizer, streaming parser, MIT license.
• DeepSeek-V4.1-Flash-1017 model card — default checkpoint, MIT weights, context window, benchmark table.
• DeepSeek-V4.1-Flash technical report — 552B backbone + 196B Engram, 8B/16B activation, 890 bytes/token KV, 45T training tokens.
• DeepSeek Zhihu post: V4.1 Flash large-scale inference deployment — Chinese production architecture write-up, 3FS, EPLB, DSpark, scheduling.
• DeepGEMM-Ascend and DeepEP-Ascend — Ascend groundwork, 2026-09-30; performance figures are vendor-run, with PoC hardware caveats.
• The Path to Open-Sourcing the DeepSeek Inference Engine — April 2025 explanation of why the full engine was not released then.
• DeepSpec — MIT-licensed training and evaluation codebase for speculative-decoding draft models (DSpark, DFlash, Eagle3).

B. Official claims we could not independently verify

These are DeepSeek's own production numbers. We quote them with “DeepSeek says” and do not treat them as benchmark results.

• 155 trillion tokens/day per cluster; US$38 revenue per GPU-hour at valley pricing. From the official Q&A and Zhihu post. The cluster size, GPU mix, and measurement window are not disclosed. Revenue is not profit.
• Load balancing reaches roughly 80% of a perfectly balanced offline benchmark. From the official Q&A.
• DSpark variable-length verification: +4–7% at the production point, up to +11% at 96 concurrent sequences per rank. From the Q&A table.
• Open-source FlashMLA is 5–10% slower on the decode kernel, at most 5% end-to-end. DeepSeek's comparison with its internal version.
• 3FS KV cache: ~100 PiB, 3-day agent TTL, 7-day chatbot TTL, ~14.4 TB logical KV per GPU, ~400 Gbps per 8 GPUs. Production design, not included in the release.

C. Secondary coverage

• CryptoBriefing, 2026-10-20 09:35 UTC — launch coverage. It additionally reports 96% cache hit, ~400 ms TTFT, 230 tokens/s (300 for coding). Those figures are not in the official README/Q&A text; we treat them as secondary and unverified.
• Pasquale Pillitteri, 2026-10-20 10:54 UTC — detailed independent analysis, chart readout, missing-component list.
• IT之家, 2026-10-20 and 第一财经, 2026-10-20 — Chinese official-statement coverage.
• Blockchain.News, TechFlow, KuCoin — short launch briefs, mostly restating the X thread.
• vLLM blog, 2026-10-07 — a separate V4.1-Flash optimization path; not a DeepInfer comparison.
• SGLang blog — a separate V4.1-Flash kernel path; not a DeepInfer comparison.

D. Community observations

• Hacker News discussion, 2026-10-20 — about 120 comments within two hours. The most recurring technical objection was the missing KV cache. Community estimates of unit economics are speculative; we do not repeat them as facts.
• DeepSeek official X thread — four posts, 2026-10-20 09:30 UTC.
• LightSeek Foundation — says TokenSpeed will learn from and adopt the optimizations.
• PR #47 (SM90/H20 RFC) and PR #46 (Ascend 910C RFC) — open proposals, not merged support.

Comparison sources

• vLLM README — PagedAttention, continuous batching, prefix caching, OpenAI-compatible API, model and hardware coverage.
• SGLang V4.1-Flash cookbook — preview image, DSpark, PD disaggregation, HiCache, router, and hardware.
• SGLang day-0 post — feature and hardware coverage for V4.1-Flash.
• TensorRT-LLM README — kernels, runtime, prefill-decode disaggregation, Wide EP, Triton/Dynamo, license.
• TensorRT-LLM support matrix — check model and GPU support before committing.

Name-disambiguation sources

• Medical DeepInfer GitHub organization (created 2016) and Slicer-DeepInfer.
• “DeepInfer: open-source deep learning deployment toolkit for image-guided therapy”, SPIE Medical Imaging 2017.
• “DeepInfer: Deep Type Inference from Smart Contract Bytecode”, ESEC/FSE 2023; code.
• DeepInfra — commercial inference provider; different spelling, different entity.

What we could not read

The Chinese Zhihu post contains a “线上系统实际统计” image. Our environment could not load the image host (SSL error), and the text mirror only returned the article text. The 96% cache hit, ~400 ms TTFT, and 230 tok/s figures attributed to that dashboard therefore remain secondary. We do not put them in headlines, spec strips, or benchmark tables.

We also did not have access to SM100 or Ascend 950 hardware, so we did not run the code, reproduce the charts, or measure Ascend performance. Every performance statement on this site is a source reading, not a hands-on result.

How to cite this site

If you cite a factual claim, cite the primary source linked on the page rather than this dossier. If you cite our analysis or the name-disambiguation table, link to the specific page and include the access date:

DeepInfer Field Notes, “Which DeepInfer? A name collision, resolved,” accessed 20 October 2026, https://deepinfer.net/name/.

Back to About, or read the limits.