DeepSeek hedges on its own published specs in self-interview

Manish Shahi, a software engineer who writes at manish.sh, ran a single chat interview with the DeepSeek model about how it works, exported 21 July 2026, as the second entry in his "Inside LLMs" series after an earlier post on Kimi K2.6. He then checked the model's answers against DeepSeek's public papers. Asked what it actually knows about itself versus what it is only guessing, DeepSeek split its answer into three categories on its own, before he asked for that framing: observation, inference, and guess. It listed public DeepSeek innovations it knows from papers rather than from reading its own weights, among them MLA for compressing the KV cache, DeepSeekMoE's shared and routed experts, auxiliary-loss-free load balancing, multi-token prediction as a training objective, and FP8 mixed-precision training. But on the specific numbers, it hedged: it did not state the 256 experts and 671B total, 37B active parameters that DeepSeek-V3's own documentation gives plainly. It offered "~1M tokens" as an educated guess about what earlier versions had announced for context length, then relabeled that as a low-confidence assumption rather than a fact about the model answering the questions. The model self-reported that it is the latest version, with a knowledge cutoff of May 2025, and said it is not labeled a reasoning model in its system prompt, unlike DeepSeek-R1, inferring from that rather than confirming that it is a general-purpose instruct/chat model. Asked what parts of its own implementation are hidden from it, DeepSeek listed a long permanent blind spot: its weights, layer count, hidden size, total parameter count, per-token expert routing, and pre-sampling logits; its tokenizer vocabulary and exact token boundaries; hidden states and attention maps between tokens; its training documents and full RLHF or DPO recipe; its GPU count and serving tricks such as beam search, best-of-N, or speculative decoding; and even whether its current answer is correct, since "let me double-check" is just more generated tokens from the same engine. On context handling, it inferred an eviction order for what gets dropped from the window when it fills: oldest messages first, then middle, then recent-but-not-immediate, then the current exchange, with the system prompt never evicted; it illustrated this with an example of forgetting a cat named Luna if the message that introduced the name is evicted. On generation, it described having no inspectable draft or plan before the first output token, only an implicit direction in activation space, and walked through prefill (processing the whole input in one pass to build the KV cache) versus decode (one sequential forward pass per output token). On hidden reasoning, it drew a distinction between two different things people call "hidden thinking": continuous latent activations present in every chat model's forward pass, which cannot be read as text even by the model itself, and separate reasoning tokens of the kind DeepSeek-R1 generates before its visible answer. It said it does not know whether this particular instance produces the second kind, since it is not labeled R1 and no reasoning tokens appeared in its own context, though they could in theory have been stripped out before it saw them. On tools and memory, which were not active during the interview, it described the general pattern rather than demonstrating it live: a product-level memory system retrieves old facts and pastes them into the prompt before the model reads them, and a tool call pauses generation while infrastructure runs the tool and a new forward pass continues. On hallucination, it attributed the root cause to being a token predictor rather than a knowledge-lookup system, citing pattern learning without guaranteed biographical accuracy, pressure to continue the sequence, blending concepts into fluent fiction, and hyper-specific or obscure requests as factors that invite invention, with exact dates and numbers named as a particular weak point. The captured text ends mid-sentence at that point; the planned later sections on MoE and MLA in plain terms, long-context attention, and a fuller comparison of the chat's claims against the published papers are not part of what was retrieved, so beyond the single hedge on the 256-expert, 671B/37B-parameter specs, no further tally of matches or mismatches between the interview and the papers is available.

Key facts

Manish Shahi interviewed the DeepSeek chat model about its own architecture and reasoning in a single chat exported 21 July 2026, the second post in his "Inside LLMs" series after Kimi K2.6.
DeepSeek hedged on specifics its own documentation states plainly: it did not give the 256 experts and 671B total/37B active parameters that DeepSeek-V3's public docs list.
DeepSeek said it is the latest version with a knowledge cutoff of May 2025, not labeled DeepSeek-R1, and listed a long set of things permanently hidden from it, including its own weights, parameter counts, training data and full RLHF recipe.
It distinguished two kinds of "hidden reasoning": latent activations present in every chat model versus separate visible reasoning tokens like DeepSeek-R1 generates, and said it does not know which applies to itself.
The captured text cuts off mid-sentence in the hallucinations section, before the post's planned direct comparison of the chat's claims against DeepSeek's published papers.

Why it matters

The piece is a data point on how much a chat model's account of its own internals can be trusted. DeepSeek volunteered a careful three-way split between what it observed, inferred, and guessed, and named real published DeepSeek techniques such as MLA and DeepSeekMoE from training on papers rather than from inspecting its own weights. Yet when asked for the concrete numbers those same papers give, such as expert count and parameter totals, it hedged rather than stating them. That gap between fluent self-description and accurate self-report is the core finding: a model can sound authoritative about its own architecture while getting the plain facts about that architecture wrong or omitting them.

Who it affects

Anyone who uses a chat model's own answers as a source for how that model works: developers debugging prompts, people writing explainers or documentation from chatbot self-reports, and researchers studying LLM introspection and self-knowledge.

How to use it

The post's own practical rule is to read arXiv or vendor documentation for architecture numbers, and use the chat interview only for behavior and prompting intuition, such as how context windows evict old messages or how tool calls pause and resume generation. Treat any specific figure a model states about its own parameters, context limit, or training as unverified until checked against a published source.

How solid is it

The method is a single chat session with one model instance, checked by the author against public DeepSeek papers, not a systematic or repeated test. The model itself could not confirm which exact version or deployment it was, and the retrieved text stops before the post's promised head-to-head comparison of chat claims versus paper claims is laid out in full, so only one specific hedge (the 256-expert, 671B/37B-parameter specs) is confirmed as a mismatch here.

Risks and caveats

The interview format means everything DeepSeek says about "why" it behaves a certain way is itself generated text, not a report of an observed internal process, a point the post makes explicitly. The source text available for this retelling cuts off mid-sentence before reaching the sections on MoE and MLA in plain terms, long-context attention, and the full paper-versus-chat comparison, so the extent of agreement or disagreement between DeepSeek's self-report and its published architecture beyond the one confirmed hedge is not established here.

“I am guessing”

— DeepSeek, on its own educated guesses about its architecture during the interview