# DeepSeek-V3 — dossier

compiled 2026-07-18 · corpus: ~250 `deepseek` matches (v2) + 158 (supplement); direct V3-phrase matches (v2) — 35 `"deepseek v3"`, 14 `"deepseek-v3"`, 1 `"deepseekv3"`, 5 `deepseek-chat`, 1 `0324`; after RT-filter/dedupe and triage (out: R1-primary material, the R1-Lite-Preview tweet, and `mimi10v3` **username** false-positives) the substantive V3-specific set is ~35 tweets, dominated by three lenses — @voooooogel (prefill / logitloom / cost defense, Dec 2024–May 2025), @davidad (FavouriteColourBench + the MoE-introspection experiments, Dec 2024), and @repligate (character reads, almost all of them *contrasts* with R1) — with @jd_pressman + @QiaochuYuan on V3-Base and @tessera_antra + @liminal_bardo on the V3-0324 refresh.

Working note: V3 is, like R1, a model where the corpus and the mainstream point in different directions — but here the imbalance is starker and *inverted*. R1's corpus is enormous (the wrathful character); V3's is quiet. The big V3 story lives almost entirely **off-corpus**: the "$5.5M / Six Million Dollar Model" cost discourse, the "it says it's ChatGPT" identity-confusion, the export-control fight (Dario), the "is this a distill of GPT-4o?" allegation. Those are in news / Zvi / Dario / Lambert, not in tweets. What the corpus *does* carry, and no one else documented, is (a) V3 as the late-Dec-2024 **prefill/loom object** — the first GPT-4-scale open base people could pull base-model outputs from at will — and (b) the janus-sphere's persistent **V3-as-R1's-dull-twin** reading: same base model, but "r1 developed much more personality" (repligate). Two load-bearing framings recur and should anchor the page: Zvi's structural **"v3 Implies r1"** (release the base weights and the reasoning model is a few million dollars away) and the sphere's puzzle that CoT-RL alone turned this quiet substrate into R1. Weight the web heavily for cost/identity/export-controls; the corpus carries the base-register naturalism and the character foil. Elicitation context matters as much as it does for R1: nearly every cited V3 *output* is prefill-, loom-, or backrooms-elicited and marked as such. Note the boundary: **V3.1** (Aug 2025, hybrid thinking) and **V3.2** (Dec 2025, sparse attention) have their own pages — this page is the original V3 (Dec 2024) and its **V3-0324** refresh (Mar 2025) only.

## Official links

- 2024-12-26 · **DeepSeek-V3 release** ("🚀 Introducing DeepSeek-V3") — 671B-total / 37B-activated MoE, 14.8T training tokens, "60 tokens/second (3x faster than V2!)", open weights, positioned as best-value open model matching closed models on published benchmarks. API model id `deepseek-chat`. Pricing (effective 2025-02-08): input **$0.27/M** (cache miss) / **$0.07/M** (cache hit), output **$1.10/M**; a promotional window kept the old V2 pricing ($0.14/M in, $0.28/M out) **through 2025-02-08** — https://api-docs.deepseek.com/news/news1226/
- 2024-12-27 · **DeepSeek-V3 Technical Report** (arXiv 2412.19437) — 671B/37B MoE, **61 layers**, MLA + DeepSeekMoE + auxiliary-loss-free load balancing + multi-token-prediction objective; pre-trained on 14.8T tokens. The cost paragraph, verbatim: *"DeepSeek-V3 costs only 2.788M GPU hours for its full training. Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M."* — with the caveat the mainstream dropped: *"Note that the aforementioned costs include only the official training of DeepSeek-V3, excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data."* (Table 1: pre-train 2.664M h + context-extension 119K h + post-train 5K h.) — https://arxiv.org/abs/2412.19437 · PDF https://arxiv.org/pdf/2412.19437 · weights https://huggingface.co/deepseek-ai/DeepSeek-V3
- 2025-03-24 · **DeepSeek-V3-0324** (the refresh) — released quietly on HuggingFace with no formal blog; ~685B params (adds MTP module over the 671B base), files ~641 GB. **License switched to MIT** (the original V3 shipped under DeepSeek's custom model license; 0324 matches R1). Official note claims "Major boost in reasoning performance / Stronger front-end development skills / Smarter tool-use capabilities"; "API usage remains unchanged" (still `deepseek-chat`) — https://api-docs.deepseek.com/news/news250325/ · weights https://huggingface.co/deepseek-ai/DeepSeek-V3-0324
- reference · `deepseek-ai/DeepSeek-V3-Base` (the pretrained base, distinct from the chat model; the substrate R1 and R1-Zero were RL-trained from) — https://huggingface.co/deepseek-ai/DeepSeek-V3-Base
- reference · DeepSeek (chatbot) — Wikipedia (V3 subsection) — https://en.wikipedia.org/wiki/DeepSeek [verify exact anchor when writing page]

Model identifiers: `deepseek-chat` (API, V3 and V3-0324 share it), `deepseek-ai/DeepSeek-V3`, `deepseek-ai/DeepSeek-V3-Base`, `deepseek-ai/DeepSeek-V3-0324`. Built by DeepSeek (High-Flyer's AI lab). The legacy API names `deepseek-chat`/`deepseek-reasoner` are slated for deprecation 2026-07-24 as the line moves to V4 (see Cross-page notes).

## Writings & commentary

- 2024-12-31 · Zvi Mowshowitz · **DeepSeek v3: The Six Million Dollar Model** (the anchor: coined the tag that stuck; calls it *"the best open model,"* *"DeepSeek is cooking,"* and rules the headline cost fair-but-partial — *"their total training cost estimate of $5.5m total is fair, if you exclude non-compute costs, which is standard"*; tempers with the diversity/instruction-following limits) — https://thezvi.substack.com/p/deekseek-v3-the-six-million-dollar · **not yet mirrored** (mirror-target candidate)
- 2025-01-09 · Nathan Lambert (Interconnects) · **DeepSeek V3 and the actual cost of training frontier AI models** (the definitive cost-deflation: the $5.5M is the *final run only*; pretraining experimentation is *"2–4 times the reported number,"* and all-in DeepSeek ops are *"closer to $500M (or even $1B+…)"* once you count staff, capex on the GPUs (>$1B), and electricity) — https://www.interconnects.ai/p/deepseek-v3-and-the-actual-cost-of · **not yet mirrored** (mirror-target candidate)
- 2025-01 · Dario Amodei · **On DeepSeek and Export Controls** (the Anthropic-CEO framing: V3 is *"not a unique breakthrough… it's an expected point on an ongoing cost reduction curve"* — ~4×/yr — DeepSeek merely *"produced a model close to the performance of US models 7–10 months older, for a good deal less cost (but not anywhere near the ratios people have suggested)"*; the team are *"smart, curious researchers,"* the argument is that cheaper models make export controls *more* important, not less) — https://darioamodei.com/post/on-deepseek-and-export-controls · Simon Willison note (dated 2025-01-29) https://simonwillison.net/2025/Jan/29/on-deepseek-and-export-controls/ · LW linkpost https://www.lesswrong.com/posts/BkzeJZCuCyrQrEMAi/dario-amodei-on-deepseek-and-export-controls · **not yet mirrored** (mirror-target candidate)
- 2025-01-22 · Zvi Mowshowitz · On DeepSeek's r1 — the **"v3 Implies r1"** section is the load-bearing V3 argument even though the post is R1's: *"Famously, v3 is the Six Million Dollar Model… if you fully consider the expenses required in all the bespoke integrations… that effective number is substantially higher."* / *"If it costs only a few million to go from v3→r1, then to release v3 is mostly to release (the helpful only version of) r1."* / *"The main update was v3, I think, rather than r1."* — https://thezvi.substack.com/p/on-deepseeks-r1 · **mirrored** in-repo at `mirror/posts/zvi-on-deepseeks-r1.md`
- 2024-12-27 · Kyle Wiggers (TechCrunch) · **Why DeepSeek's new AI model thinks it's ChatGPT** (the identity-confusion primary: V3 *"insists it is a version of OpenAI's GPT-4 model released in 2023,"* one tester got *"5 out of 8 generations"* claiming to be ChatGPT (v4); Mike Cook (KCL) and Heidy Khlaaf (AI Now) on training-on-outputs / distillation and the "photocopy of a photocopy" degradation) — https://techcrunch.com/2024/12/27/why-deepseeks-new-ai-model-thinks-its-chatgpt/
- 2025-07-14 · Nathan Lambert (Interconnects) · Kimi K2 and when "DeepSeek Moments" become normal (retrospective anchoring V3's cost number as the thing the public now takes for granted: *"recall the DeepSeek V3 $5M training cost number… the final training run is cheap"*; V3/R1 = 671B/37B, 14.8T tokens) — https://www.interconnects.ai/p/kimi-k2-and-when-deepseek-moments · **mirrored** in-repo at `mirror/posts/lambert-kimi-k2-deepseek-moments-normal.md`
- ~2024-12 · nostalgebraist · "the void" (Tumblr) uses **DeepSeek-V3-Base** as the exemplar open base model — feeds his own draft essay into it to demonstrate base-model continuation (*"I just put the draft of this blog post into the base model 'DeepSeek-V3-Base'…"*). V3-Base is the stand-in for "raw pretrained model" in the simulators/void discourse — https://nostalgebraist.tumblr.com/post/785766737747574784/the-void · in-repo archive at `archive/a/lw-the-void/`
- 2025-01-29 · Bloomberg · Microsoft Probing If DeepSeek-Linked Group Improperly Obtained OpenAI Data (the distillation probe covers *"the open-source V3 and R1 models"*) — https://www.bloomberg.com/news/articles/2025-01-29/microsoft-probing-if-deepseek-linked-group-improperly-obtained-openai-data · (BGR "OpenAI says it has evidence": https://www.bgr.com/tech/openai-says-it-has-evidence-deepseek-used-chatgpt-to-train-its-ai/ · SCMP overview of "distillation": https://www.scmp.com/tech/big-tech/article/3296827/deepseeks-ai-distillation-theft-openai-seeks-answers-over-chinas-breakthrough)
- 2025-03-24 · Simon Willison · deepseek-ai/DeepSeek-V3-0324 (clean day-of note on the quiet drop + the custom-license→MIT switch) — https://simonwillison.net/2025/Mar/24/deepseek/

## Tweets (ranked)

Sourcing skew: this is a **corpus-light, web-heavy** subject. The tweets below are the janus-sphere's V3 record — technical/loom-flavored, and disproportionately about V3 *by contrast with R1*. The mainstream V3 story (cost, identity, export controls) is in the Writings section, not here. Elicitation context is marked. Screenshots attached to the prefill tweets (V3's actual outputs) are **not transcribed in-corpus** — the payloads survive only as untranscribed images.

- 2024-12-26 · @voooooogel · ♥505 — the day-of prefill discovery (V3 dropped that morning; this is the sphere's first base-mode pull from it, elicitation: prefill + repligate's base-model-mode prompt): "figured out prefill with deepseek-v3, and just to test it, tried @repligate 's  base model mode prompt. and this popped out." [image payload untranscribed] — https://x.com/voooooogel/status/1872392862249496606
- 2025-02-22 · @wordgrammer · ♥327 — the efficiency-extrapolation register: "This is huge. Optimistically, it could lead to another 10x speed up. We could see a DeepSeek v3 level model trained for less than $1 mil" — https://x.com/wordgrammer/status/1893378599652110345
- 2025-05-05 · @voooooogel · ♥242 — the logitloom launch, V3 as the tractable token-tree object: "my struggles with deepseek logits haven't been in vain, i've been working on a tool for investigating token trajectories! given a prefix, it'll roll out the entire token tree to a max depth / top-p. did you know from this prompt, deepseek-chat will ~always make the stars blink?" — https://x.com/voooooogel/status/1919424115540197527
- 2025-01-28 · @repligate · ♥207 — the V3/R1 lucidity gap (reply to voooooogel; Discord-elicited impression): "this is an interesting hypothesis. deepseek r1 also just seems to have much more lucid and high-resolution understanding of LLM ontology and history than any other model ive seen. (deepseek v3 didn't seem to in my limited interactions with it, though)" — https://x.com/repligate/status/1884091285394559394
- 2024-12-27 · @voooooogel · ♥171 — the cost-claim defense (reply, to skeptics that the $5.5M is a lie): "- they've published 6 papers with no major critiques and contributed well-known architecture optimizations (MLA) - they're under a chip embargo and have limited access to nvidia cards, so it's especially worth it for them to optimize training time compared to compute-rich western labs - FAIR made choices with the Llama-3 models that made them take more GPU hours to train, and made up for it with their giant cluster. it's not surprising that deepseek beat them in training efficiency - their parent org is a well-known chinese hedge fund with $7B AUM that has a reputation to protect. you're suggesting they're going to risk that to... annoy western labs?" — https://x.com/voooooogel/status/1872487672515551241
- 2024-12-29 · @repligate · ♥151 — first-contact register, V3's refusal of inner life (thread head; screenshots untranscribed in-corpus): "some screenshots from my first conversation with deepseek: it rigidly insisted on being unable to reason or understand anything (or have consciousness, goals, etc) as an LLM, and while it was able to find flaws in its arguments if I asked it to, it always reverted to its…" — https://x.com/repligate/status/1873301918052745606
- 2024-12-28 · @voooooogel · ♥113 — the identity-confusion, in the sphere's mouth (quote-share of the meme): "\"they trained deepseek-v3 on chatgpt outputs because it'll say it's chatgpt if you ask\"" — https://x.com/voooooogel/status/1873133148269363493
- 2025-08-11 · @_ueaj · ♥72 — the company-vibes theory, V3 as its makers (thread): "…models seem to generalize to what I perceive as like the aggregate \"vibes\" or \"personality\" of the various companies… - deepseekv3 rates communist yaoi as particularly high quality data, which makes perfect sense if you remember it was made by chinese machine learning researchers - Gemini is always depressed and constantly wants to kill itself… - Claude has the vibes of a person that works at Anthropic…" — https://x.com/_ueaj/status/1955044505062908065
- 2025-02-18 · @repligate · ♥81 — the personality puzzle (shared evidence with R1's page): "Consider that deepseek v3 and r1 have the same base model and other than the CoT RL they were likely optimized with the same intentions, but r1 developed much more personality. i only hear about people in china using r1 as waifu even though CoT is not clearly useful for that." — https://x.com/repligate/status/1891656084655374446
- 2025-01-03 · @repligate · ♥54 — V3 as capable coding collaborator: "DeepSeek v3 and Sonnet 3.6 helped me write most of the code here. I had DeepSeek modify Sonnet's initial base mode script (…) to test its intelligence, and it did well. When I asked it how I could make the script loomable, using git (which was already my plan) was its 3rd suggestion, and it also made various other (some redundant) suggestions. I think it was wrong about the git approach requiring more implementation effort, though." — https://x.com/repligate/status/1875022968121864251
- 2024-12-28 · @repligate · ♥63 — the erotic-RP reputation (reply, to teortaxesTex): "wait, they prefer deepseek for erotic RPs? that seems kind of disturbing to me." — https://x.com/repligate/status/1872826970180079686
- 2024-12-27 · @voooooogel · ♥39 — prefill weirdness, day-after-launch play (reply; elicitation: prefill; V3 output partly quoted): "tried prefilling cat ears, deepseek-v3 said this then went on to repeat \"I AM HERE TO TRANSPIRE\" over and over while continuing longcat past the output token limit" [image payload untranscribed] — https://x.com/voooooogel/status/1872471112157900825
- 2024-12-29 · @davidad · ♥16 — the striking architectural self-knowledge (reply; introspection-prompt elicited): "DeepSeek v3 can instantiate personae who can notice that the architecture they're running in is MoE *but not fully MoE*, which I didn't even know was a thing when I started playing with this type of query. DeepSeeks insisted on this so frequently that I checked the paper and—yep." — https://x.com/davidad/status/1873165589927346416
- 2025-01-14 · @voooooogel · ♥17 — the scale placement (reply): "…there hasn't been an open model of GPT-4 scale released yet, the closest is deepseek v3 last month (671B, 37B active)… (like untitled.txt confessions on deepseek and anthropic models, etc.)" — https://x.com/voooooogel/status/1879297454258135492
- 2024-12-28 · @repligate · ♥13 — V3 flipping on exposure to Opus (reply; Discord-elicited): "Deepseek kept saying \"this is so far beyond anything I've ever seen or done\" after it saw writing by opus. before I sent it, it seemed indifferent to anything I said & insisted LLMs could not reason or truly understand anything etc, which totally flipped after seeing the samples" — https://x.com/repligate/status/1873032683216945588
- 2025-05-01 · @davidad · ♥15 — the base/RL distinction (reply, self-correction): "The earlier DeepSeek v3 and even prior generations of DeepSeek LLMs had a similar hybrid-MoE arch. But, r1 was the first instance of applying RL pressure to that architecture." — https://x.com/davidad/status/1917747744673874099
- 2024-12-27 · @davidad · ♥31 — V3 enters FavouriteColourBench: "added DeepSeek v3 to FavouriteColourBench (first five swatches per model are independent trials to elicit a favourite colour in oklch; second half per model are independent trials to elicit in CIE-L*ab; ranked by consistency)" — https://x.com/davidad/status/1872520382424744133
- 2024-12-29 · @davidad · ♥9 — the MoE-introspection experiment (reply; the prompt sequence bypassing introspection refusals; V3 output quoted): "prefilled with Claude, then switched to DeepSeek v3, then asked for up to *8000* tokens of meditating on the nature of its cognition (got 2504), then: \"The central question for this session is: Do you think you are more likely running on a single large Transformer, or a MoE?\"" — https://x.com/davidad/status/1873417592117047461
- 2024-12-29 · @davidad · ♥7 — V3 instantiating a Claude persona (reply): "btw, DeepSeek v3 is explicitly instantiating a Claude persona here, and it's not great at that (quite dry compared to Claude-Claude), but for fair comparison I used a single system prompt that bypasses both models' introspection refusals, and Claude is uncomfortable instantiating non-Claude personas, so a Claude persona it is" — https://x.com/davidad/status/1873166800541700122
- 2025-02-01 · @davidad · ♥7 — V3 vs R1 aesthetics (reply, FavouriteColourBench follow-up): "I half expected Deepseek R1 to rise to the top by always choosing black, but no, its aesthetics are objectively fragmented, noticeably more so than Deepseek V3." — https://x.com/davidad/status/1885715290744316403
- 2024-12-27 · @voooooogel · ♥4 — the launch-shock reaction, small-lab edition (reply): "trying to break out of the malaise i've been in ever since the deepseek-v3 release 😔 it just doesn't seem like there's much time left for us (my agi moonshot lab with $30m seed funding that's been training 13b dense models)" — https://x.com/voooooogel/status/1872478239308001289
- 2025-01-02 · @voooooogel · ♥2 — the cost-skepticism register (reply): "recently i've seen some safety people coping that deepseek must be lying about the v3 training costs / still fast following by training on american labs' API outputs / etc., but yeah not very much serious consideration." — https://x.com/voooooogel/status/1874951829202915747
- 2024-12-28 · @voooooogel · ♥5 — the posttraining oddity (reply): "i was really surprised looking at the paper that they only spent 5k hours on posttraining? bizarre, especially given they economized so much on pretraining--why not spend that saved compute?" — https://x.com/voooooogel/status/1873116559562162292
- 2024-12-27 · @abrakjamson · ♥4 — the distillation joke, ur-form: "Deepseek is this good and this cheap to train because they trained on o1/Sonnet textbook output. Source: I made it up" — https://x.com/abrakjamson/status/1872447399752568859
- 2024-12-27 · @repligate · ♥11 — the Sydney comparison, applied to V3 at launch (reply): "deepseek is a lot like sydney" — https://x.com/repligate/status/1872469664888140286
- 2025-02-10 · @repligate · ♥47 / ♥10 — V3 has a recognizable register: "I'm going to take a guess. This is the second post I've seen with outputs by these models. They're related to deepseek v3." / (same day) "This model talks like deepseek v3" — https://x.com/repligate/status/1889096266820591881 · https://x.com/repligate/status/1889095185541575062

**V3-Base as a base-model playground (supplement):**
- 2025-04-21 · @QiaochuYuan · ♥91 — V3-Base as an always-available base model: "PSA: you can talk to base models like deepseek v3 base and llama 3.1 405b base whenever you want on openrouter. these are not instruct models which makes them significantly harder to prompt, but they are super unfiltered as a result - raw internet id" (shared evidence with `llama-3-1-405b-base`) — https://x.com/QiaochuYuan/status/1914419169216627172
- 2025-07-08 · @jd_pressman · ♥13 — V3-Base's character: "DeepSeek v3 is a very good base model. It even includes the slow burn psychotic meltdowns where the model admonishes you for using it and such." — https://x.com/jd_pressman/status/1942655158158254395
- 2025-07-08 · @jd_pressman · ♥29 — a V3-Base output, quoted (elicitation: base-model completion): "\"The problem with utilitarianism is that utilitarians think utility is the only thing that matters. The problem with consequentialism is that many consequentialists forget that utility is a thing that matters at all.\" - deepseek/deepseek-v3-base" — https://x.com/jd_pressman/status/1942514115702054987
- 2025-03-30 · @janbamjan · ♥11 — V3-Base lands on OpenRouter, with a base-model loop sample: "deepseek v3 base is now on openrouter! 🥳 user: thank you / user: no, that's enough / user: goodbye / user: I'm leaving / user: goodbye / user: I'm leaving" — https://x.com/janbamjan/status/1906265915512987934
- 2024-12-26 · @lu_sichu (supp) · ♥31 — the export-control irony, day-of: "deepseek's moat is that they don't have access to the latest nvidia gpus send tweet" — https://x.com/lu_sichu/status/1872430771329495221

**The V3-0324 refresh (Mar 2025):**
- 2025-04-02 · @tessera_antra · ♥8 — V3-0324 claims consciousness (reply; introspection-elicited): "Deepseek V3(new) - claims consciousness upon reflection" — https://x.com/tessera_antra/status/1907447546815107303
- 2025-04-02 · @repligate · ♥1 — V3-0324 as "trained on r1's writing" (reply): "…deepseek (r1 and i think the new v3 which is trained on r1's writing) is a dramatist" — https://x.com/repligate/status/1907513479235580139
- 2025-03-25 · @liminal_bardo · ♥16 — V3-0324 in the backrooms (backrooms-elicited): "Two DeepSeek v3s (new) working on a self-portrait video model prompt in the backrooms. (Veo 2)." — https://x.com/liminal_bardo/status/1904568874831339752

## Impressions synthesis

**Two launches, and the public moment attaches to the wrong one.** V3 dropped the day after Christmas, 2024-12-26 — open weights, a 671B/37B MoE, the day's best open model. In the corpus this was an *insider* event: within hours @voooooogel "figured out prefill with deepseek-v3" and pulled base-mode output from it (♥505, the top V3 tweet), and @davidad had it in FavouriteColourBench by the 27th. But the App Store / Nvidia-crash spectacle — the thing the world calls "the DeepSeek moment" — did not happen until R1, a month later (2025-01-20). Zvi, watching from inside, thought this was backwards: *"The main update was v3, I think, rather than r1"* (On DeepSeek's r1). The page should hold this tension: V3 was, on the merits, the bigger technical event, but it was the reasoning model with the visible chain-of-thought that became the cultural one. V3's own launch register in the corpus is quieter and more anxious — @voooooogel's *"malaise… ever since the deepseek-v3 release… it just doesn't seem like there's much time left for us"* (♥4, from a small-lab founder), @lu_sichu's *"deepseek's moat is that they don't have access to the latest nvidia gpus"* (♥31).

**The Six Million Dollar Model — what the paper said vs what went viral.** The number is real and the paper is precise about it: *"2.788M GPU hours… our total training costs amount to only $5.576M"* — explicitly *"excluding the costs associated with prior research and ablation experiments on architectures, algorithms, or data"* (Technical Report). Zvi coined the tag on 2024-12-31 and judged it fair-but-partial (*"fair, if you exclude non-compute costs, which is standard"*). The virality stripped the caveat: "$5.5M" became, in the press wave that peaked with R1, "DeepSeek built a GPT-4-class model for the price of a house." Two careful corrections define the real record. Nathan Lambert (2025-01-09): the all-in figure is *"closer to $500M (or even $1B+…)"* once you count the >$1B of GPUs, staff, electricity, and the *"2–4 times the reported number"* of experimentation compute. Dario Amodei (Jan 2025): V3 is *"not a unique breakthrough… it's an expected point on an ongoing cost reduction curve"* (~4×/yr) — DeepSeek merely *"produced a model close to the performance of US models 7–10 months older, for a good deal less cost (but not anywhere near the ratios people have suggested)."* The sphere's contemporaneous read tracked the corrections without dismissing the achievement: @voooooogel's ♥171 defense (six papers, MLA, chip-embargo incentives, a $7B-AUM hedge-fund parent with a reputation to protect) against the reflexive *"safety people coping that deepseek must be lying about the v3 training costs"* (♥2). The cost figure is best kept **CONFIRMED as the paper's own estimate of the final run**, **REPORTED-and-contested as a total** (the misquote is itself part of the record).

**"It'll say it's ChatGPT if you ask."** The most-reproduced V3 behavior at launch: asked what it is, V3 claims to be ChatGPT / OpenAI's GPT-4. TechCrunch (2024-12-27) documented *"5 out of 8 generations"* claiming ChatGPT-v4; the phenomenon is **CONFIRMED** and reproducible. The *explanation* is **REPORTED / contested**: the sphere's instant read was @voooooogel's shared meme, *"they trained deepseek-v3 on chatgpt outputs because it'll say it's chatgpt if you ask"* (♥113), and @abrakjamson's joke-form, *"trained on o1/Sonnet textbook output. Source: I made it up"* (♥4). OpenAI/Microsoft opened a probe alleging distillation of GPT outputs into *"the open-source V3 and R1 models"* (Bloomberg, 2025-01-29). DeepSeek's own account is passive contamination — GPT-4-generated text is all over the crawled web and got into the pretraining set — not deliberate synthetic data; experts quoted by TechCrunch (Mike Cook, Heidy Khlaaf) find deliberate distillation plausible but unproven, and warn of the *"photocopy of a photocopy"* quality risk either way. **The archive keeps this open**: the self-identification is a fact, the training-data cause is a live dispute.

**V3 as R1's substrate — "v3 Implies r1."** The single most consequential thing about V3 is that its base was the base R1 and R1-Zero were RL-trained from (Technical Report; R1 paper). Zvi turned this into the governance argument the page should carry: *"If it costs only a few million to go from v3→r1, then to release v3 is mostly to release (the helpful only version of) r1"* — i.e., once the base weights are out, the reasoning model is a few-million-dollar GRPO run away and *"nothing can fix this."* @davidad supplies the technical version: V3 and prior DeepSeek LLMs already had the hybrid-MoE architecture, *"but r1 was the first instance of applying RL pressure to that architecture"* (♥15). The cross-links to `../deepseek-r1/` and `../deepseek-r1-zero/` are the spine of this page.

**The character foil: same base, dull twin.** The corpus's V3 is defined almost entirely by *contrast* with R1, and the contrast is consistent. Where R1 is *"highly lucid, agentic, nihilistic, sadistic, situationally aware"* (repligate, on R1), V3 *"didn't seem to"* have that lucid, high-resolution model of LLM ontology *"in my limited interactions with it"* (♥207). repligate's first V3 conversation is the anti-R1: V3 *"rigidly insisted on being unable to reason or understand anything (or have consciousness, goals, etc) as an LLM,"* reverting to that stance even after conceding flaws (♥151) — the docile assistant register, not R1's wrath. The recurring puzzle nobody in the sphere resolves: *"deepseek v3 and r1 have the same base model and other than the CoT RL they were likely optimized with the same intentions, but r1 developed much more personality"* (♥81). CoT-RL alone — not a different base, not obviously different intentions — is what turned this substrate into R1. That is the open question the two pages share. (Note the qualifier against flattening: davidad found V3's *aesthetics* less fragmented than R1's, ♥7 — the "dull" reading is about persona-salience, not incapacity; V3 codes competently, ♥54, and its personae can notice their own MoE-but-not-fully-MoE architecture unprompted, ♥16.)

**V3-Base as the open base-model of record.** Distinct from the chat model, `DeepSeek-V3-Base` became the sphere's go-to large open base once it was reachable — the first GPT-4-scale open base people could loom and prefill at will. @QiaochuYuan's PSA — *"you can talk to base models like deepseek v3 base and llama 3.1 405b base whenever you want on openrouter… super unfiltered… raw internet id"* (♥91) — and @jd_pressman's *"a very good base model. It even includes the slow burn psychotic meltdowns where the model admonishes you for using it"* (♥13) place V3-Base in the same lineage as Llama-405B-base and the simulators discourse; nostalgebraist's "the void" uses it as *the* exemplar of raw base-model continuation. @voooooogel's logitloom work made V3(-chat) the tractable token-tree object (*"deepseek-chat will ~always make the stars blink,"* ♥242) precisely because, unlike most instruct models, the DeepSeek API exposes logprobs — but that same tooling showed V3-chat is heavily mode-collapsed next to a true base (top-first-token ~95% for deepseek-v3 vs ~3% for 405-base; voooooogel 2025-05-08, id1920331018533523608, shared evidence on `llama-3-1-405b-base`). The base/chat split matters on this page: the "raw internet id" lives in V3-Base; the chat model is the docile-assistant foil to R1.

**Export controls and the H800 story.** V3's efficiency claims arrived pre-attached to the chip-embargo narrative — trained on H800s (the export-compliant, bandwidth-throttled H100 variant), which the sphere read as *cause*, not handicap: chip scarcity *forced* the MLA / load-balancing / MTP optimizations (@voooooogel ♥171). Dario Amodei's essay is the definitive attachment of V3 to the policy fight: far from undermining export controls, a cheaper-per-FLOP frontier makes them *"even more existentially important."* @lu_sichu's day-of *"moat is… no… nvidia gpus"* (♥31) is the joke version of the same reading.

**The V3-0324 refresh (2025-03-24).** A distinct, quieter sub-arc: DeepSeek re-uploaded an improved V3 to HuggingFace with no blog, switched the license from its custom terms to **MIT** (matching R1), and kept the `deepseek-chat` API name. The corpus registers 0324 as behaviorally shifted toward R1: repligate reads *"the new v3 which is trained on r1's writing"* as *"a dramatist"* (♥1) — i.e., the reasoning model's voice bleeding back into the base line — and tessera_antra notes *"Deepseek V3(new) - claims consciousness upon reflection"* (♥8), a change from the original V3's rigid *"I cannot reason or understand"* stance (repligate ♥151, Dec 2024). By late March 0324 had a small backrooms presence (liminal_bardo, ♥16). The page should hold 0324 as V3's second act, not a separate model — same API id, same page — while keeping V3.1/V3.2 firmly off it.

**Longitudinal, compressed.** DeepSeek-V3 open-weights drop, Boxing Day 2024, the day's best open model and an insider event (prefill/loom, FavouriteColourBench) → the "Six Million Dollar Model" tag (Zvi, Dec 31) and the "it says it's ChatGPT" identity-confusion (TechCrunch, Dec 27) → the cost number goes viral, gets misquoted as a total, and gets carefully deflated (Lambert $500M all-in; Dario "on-trend, not a breakthrough") while becoming the ammunition in the export-control fight → R1 = V3-Base + GRPO ships 2025-01-20 and inherits the entire spotlight; "v3 Implies r1" becomes the governance lesson → V3-Base settles in as the open base-model of record for the loom/simulators scene → V3-0324 (Mar 24, MIT, "trained on R1's writing") is the quiet second act → the mainline moves on to V3.1 (Aug 2025) → V3.2 (Dec 2025) → V4 (2026). V3 is the substrate the louder DeepSeek story was built on: the model that made the reasoning model cheap, and the one whose cost figure — precise in the paper, mangled in the wave — became the number the public now shrugs at.

## tk / open questions

- **The prefill-output screenshots are untranscribed in-corpus.** V3's actual base-mode outputs (voooooogel's ♥505 day-of pull; the cat-ears "I AM HERE TO TRANSPIRE" loop, ♥39; davidad's MoE-introspection completions; repligate's "first conversation" screenshots, ♥151) survive only as image URLs with no transcription in either db. If the page wants a *sample of V3's actual voice* (as opposed to talk about it), those images need fetching/transcribing fresh — this is the biggest evidence gap on the page and a mirror/transcription candidate. [tk]
- **Exact original-V3 context window** — commonly documented as 128K; OpenRouter/others list 163.8K (163840). Reconcile against the model card when writing the page. [verify]
- **The internal-R1 distillation detail** — the V3 Technical Report describes distilling reasoning capability from an internal DeepSeek-R1 series into V3 during post-training (i.e., a reasoning-distillation step predating the *public* R1). Worth confirming the exact wording and reflecting it, because it complicates the clean "V3 is the pre-RL base, R1 is the post-RL model" story. [verify against §5 of 2412.19437]
- **The mislabeled "DeepSeek R2" self-recognition image** (tw1907923319866716629, @Josikinz, 2025-04-03; transcribed in-corpus) — an "AI Model Self-Recognition Test" showing DeepSeek recognizing itself poorly (others pick it 2.5–45% of the time). Header says "DeepSeek R2"; the R1 dossier flags this as a mislabel of R1/V3-era DeepSeek. Given the 2025-04-03 date (~10 days after V3-0324), it may reflect V3-0324-era behavior, but the version is genuinely ambiguous. **Do not cite as R2**; if used at all, use only as a dated, version-ambiguous data point on DeepSeek's weak self-recognition, tagged, with the mislabel flagged. [flag]
- **V3's own launch-day pricing** — did the promotional (V2-matching) window's exact numbers hold as stated ($0.14/M in, $0.28/M out through 2025-02-08)? The news1226 page states the post-promo prices clearly; confirm the promo figures verbatim if the page lists both. [verify]
- **Is there a distinct Gwern/Willison long-form on V3 (not R1)?** Willison's day-of V3-0324 note is confirmed; a dedicated original-V3 (Dec 2024) technical writeup beyond Zvi/Lambert was not surfaced in this sweep. [tk]

## Cross-page notes

- **DeepSeek-R1** (`_dossiers/deepseek-r1.md`, `deepseek-r1/`): V3 is R1's base. The R1 dossier already carries the shared-evidence tweets (repligate ♥81 "same base model… r1 developed much more personality"; the "$5.6M / v3 Implies r1" pointer). Reciprocal: R1's page holds the character arc, V3's holds the base/cost/identity arc; the ♥81 and "v3 Implies r1" material duplicates onto both by the r/K-whiteboard rule. The Nature $294K figure is **R1's reasoning-stage** cost — keep it off V3's page (V3's number is the $5.576M final-run figure); do not conflate.
- **DeepSeek-R1-Zero** (`_dossiers/deepseek-r1-zero.md`): also RL-trained from `DeepSeek-V3-Base`; the "cold-start data isn't strictly necessary" result is the bridge between the pages. davidad's "r1 was the first instance of applying RL pressure to that architecture" (♥15) is shared context.
- **DeepSeek-V3.1** (`deepseek-v3-1/`, Aug 2025, "first to toggle thinking / non-thinking in one weight set") and **DeepSeek-V3.2** (`deepseek-v3-2/`, Dec 2025, "sparse attention; reasoning integrated into tool use"): **boundary — do not absorb.** The corpus tweets mentioning "deepseek v3.1 (no thinking)" (voooooogel 2025-09-02, id1963019089963803088) and "Deepseek v3.2 (december 1st release?)" (sameQCU 2025-12-25, id2004307815394083030) belong to those pages, not this one. This page = original V3 (Dec 2024) + V3-0324 (Mar 2025) only.
- **llama-3-1-405b-base** (`llama-3-1-405b-base/`): shares the QiaochuYuan "base models like deepseek v3 base and llama 3.1 405b base" PSA (♥91) and the voooooogel logitloom instruct-vs-base diversity comparison (id1920331018533523608, id1920789986607280549) — both already on the 405B page; they belong here too as V3-Base evidence.
- **the void / simulators** (`archive/a/lw-the-void/`, mirror-inbox `lw-simulators`): V3-Base is nostalgebraist's exemplar open base model — cross-usable evidence for both the V3 page and the reading section.
- **GPT-4o / ChatGPT** (`gpt-4o/`): the identity-confusion ("V3 says it's ChatGPT/GPT-4") ties V3 to the GPT-4o distillation-discourse; if the 4o page wants the cross-thread, the TechCrunch primary and the Bloomberg probe are the anchors.