On October 17, 2026, a 552-billion-parameter model appeared on Hugging Face under the MIT license with no press release, no launch livestream, and a filename instead of a product name. DeepSeek V4.1-Flash-1017 was submitted to Hacker News by a user called theykk, and about eighteen hours later the thread sat at 732 points and 602 comments. On OpenRouter's public ranking page the same day, that model was processing 4.1 trillion tokens per day — roughly 2.7 times the daily volume of the second-place model, and 124.6 trillion tokens over the trailing 30 days. The genuinely strange part is not that a Chinese lab shipped something good. It is that the weights are free, sitting in a public repository right now, and almost nobody arguing about them in that thread can actually run them.
What DeepSeek Actually Put on Hugging Face
The release is a multimodal mixture-of-experts model with a 552B-parameter backbone — 763B if you count every component shipped alongside it — that activates only 8B parameters per token during prefill and 16B per token during decode. Context length goes up to one million tokens, with the model card recommending a maximum output of at least 256K. The repository and the weights are both under the MIT license, published on Hugging Face and mirrored to ModelScope. MIT is the detail doing the heavy lifting here: not a bespoke "community license" with a user-count trigger, not a research-only clause. Download it, fine-tune it, resell access to it.
The Sparsity Is the Whole Trick
The architecture is a Causal Encoder-Decoder MoE with 40 layers, split 20 encoder and 20 decoder. Each MoE layer holds one shared expert plus 384 routed experts, of which six fire per token. That is how you get a 552B model that costs about as much to run as an 8B one at prefill. On top of that sits something DeepSeek calls Compressed Sparse Attention 2, with three static attention modes and FP4 main KV caching, which pushes the global KV cache down to 890 bytes per token. Do that arithmetic and a full million-token context costs a hair under 900MB of cache — which is the actual reason the 1M context number is a product feature rather than a benchmark stunt. Pre-training ran on 45 trillion tokens of multimodal data, post-training went supervised fine-tuning to reinforcement learning to on-policy distillation, and the released model carries a vision encoder (DeepSeek-ViT), a 196B-parameter conditional memory component, speculative decoding, and a continuously adjustable reasoning-effort dial from 1 to 100.
The Scores, With One Caveat
The model card reports 74.2% resolved on DeepSWE v1.1, 90.6% pass rate on Terminal-Bench 2.1, 90.9% on GPQA Diamond, 79.4% pass@1 on HumanEval, and 93.0% on GSM8K. At least one trade writeup published on release day headlined a 77.0 DeepSWE figure instead, so treat the exact coding number as unsettled pending a cleaner reproduction. What is more interesting than any single score is the shape of them: 90.6% on a terminal-agent benchmark against 79.4% on plain HumanEval function completion. This is a model tuned for driving a shell and a toolchain, not for winning puzzle-style code challenges — which matches what developers in the thread actually describe using it for.
"Most-Used" Is a Measurable Claim, Not a Press Line
OpenRouter rankings measure revealed preference — which models developers route real, paid traffic through — and that makes them the closest thing to a public usage census the industry has. Here is the top of the board as of October 17, 2026, with current list prices per million tokens alongside:
| Rank | Model | Provider | Tokens/day | 30-day total | Price in/out per 1M |
|---|---|---|---|---|---|
| 1 | DeepSeek V4.1 Flash | DeepSeek | 4.1T | 124.6T | $0.30 / $1.20 |
| 2 | GLM 5.3 Flash | Z.ai | 1.5T | 56.2T | $0.15 / $0.50 |
| 3 | Mimo V2.6 Flash | Xiaomi | 1.2T | 35.5T | not listed |
| 4 | Hy4 Preview | Tencent | 907.4B | 36.5T | not listed |
| 5 | Claude Opus 5.5 | Anthropic | 799.9B | 13.9T | $4.00 / $20.00 |
| 6 | GPT-6 Luna | OpenAI | 706.1B | 21.9T | $0.10 / $0.50 |
Two things jump out of that table. First, the entire top four is Chinese — DeepSeek, Z.ai, Xiaomi, Tencent — before you reach a US lab. Second, the 30-day column is less flattering to the frontier than the daily one: DeepSeek V4.1 Flash moved roughly nine times the token volume of Claude Opus 5.5 over the month. Weekly figures reported elsewhere during October ranged from about 36.8 trillion to 41 trillion tokens depending on the snapshot, with DeepSeek models accounting for somewhere around a quarter of all requests on the platform. Any one of those numbers will be stale by next week; the ratio has been stable for months.
The Price Is Doing a Lot of the Work
DeepSeek's hosted API lists $0.30 per million input tokens and $1.20 per million output, dropping to $0.15 and $0.60 off-peak. Set that against Claude Opus 5.5 at $4.00 and $20.00 and you get a 13x gap on input and nearly 17x on output at list price — call it 26x and 33x if your workload can run off-peak. For context on how unusual that is, the median frontier-model price in October 2026 sat at $3.75 per million tokens, down about 9% since June. The honest comparison, though, is not DeepSeek against Opus. It is DeepSeek against GLM 5.3 Flash and GPT-6 Luna, both of which undercut it on paper and both of which lost the traffic anyway. Price explains why people tried it. It does not explain why they stayed.
"Anyone Can Download It" Has a 510GB Asterisk
This is where the open-weights story gets honest. The full download measures 510.30GB. GGUF quantizations land at 264.5GB for Q2_K, 347.3GB for Q3_K_M, and 444.7GB for Q4_K_M — meaning even the most aggressive quantization is more than double the unified memory in the largest MacBook configuration on sale. There is no partial-offload trick that rescues a laptop here. In deployment the picture is slightly better than the download size suggests, because the Engram lookup tables can live on storage: roughly 150–200GiB of memory for the backbone plus about 189GiB of fast SSD.
What a Working Local Setup Actually Looks Like
Two configurations have been reported as working. Two linked DGX Sparks, 256GB combined, running EXL3 quantization at 2.77–3.0 bits per weight on the routed experts, deliver 25–41 tokens per second of single-stream decode and roughly 960–2,050 tokens per second of prefill — and require a RoCE-enabled QSFP cable between the units. A 256GB Mac Studio loads a Q2 build at about 152GB resident, with no published throughput numbers yet. A 512GB Mac Studio should take a Q4 build in theory; nobody has posted measurements.
The Tooling Isn't There Yet
The friction is more mundane than the memory math. llama.cpp's converter registers DeepseekV4ForCausalLM, while V4.1 declares DeepseekV41ForCausalLM — a one-identifier mismatch that blocks the entire mainline local-inference path, and with it Ollama and LM Studio. Server-side engines (vLLM, SGLang, Transformers) are supported out of the gate. In the Hacker News thread, user alfiedotwtf was already asking which quantization to prefer, Q4_K_S or Q3_XS, having noticed that larger quants loop less — a question nobody in the thread could answer with official numbers. Here is the uncomfortable summary: for everyone except a few hundred people with the right hardware, this MIT-licensed model is something you rent, not something you own.
| DeepSeek's own API | Third-party host | Local weights | |
|---|---|---|---|
| Cost | $0.30 / $1.20 per 1M, half that off-peak | Varies; typically above DeepSeek's rate | Hardware up front, electricity after |
| Where your prompts go | DeepSeek, in China | The host's jurisdiction | Nowhere |
| Zero-data-retention | Raised as a concern in the thread | Offered by some, e.g. Fireworks | Not applicable |
| Hardware needed | None | None | ~150–200GiB RAM + ~189GiB fast SSD |
| Measured speed | Thread consensus: fastest of the group | Provider-dependent | 25–41 tok/s decode on dual DGX Spark |
| Works today with Ollama/LM Studio | N/A | N/A | No — class-name mismatch |
The Loudest Objection Isn't Quality
Read 600 comments and the pattern is clear: barely anyone is arguing the model is bad. The argument is about where the tokens go. Commenter peri-cl put the practical version of it — DeepSeek's own cloud raises zero-data-retention questions, some third-party hosts such as Fireworks offer ZDR contractually, and, in the same breath, the US labs train on data from their own subsidized plans. That last clause is why the thread never resolved into a tidy East-versus-West split.
The Case for Sending Your Tokens to Hangzhou, Stated Fairly
User dannyw made the argument that got the most traction: he would rather his data go to China than to "NSA/CIA; who has far better ability to hurt me," and he credited DeepSeek specifically for still being MIT licensed. That is a threat-model argument, not a geopolitical one — whichever state can actually act against you is the one that matters. igetspam described getting "nearly equivalent results without the constraints," comparing frontier model behavior to a "chaperone at a church dance," and serf noted that Chinese models don't moralize about licensing and legal edge cases. OsrsNeedsf2P praised the raw throughput — "blazes through work" at "quite good quality" — and added a reason that has nothing to do with capability: not getting randomly blocked for cyber review. Elsewhere the thread drifted into whether the V-Dem human rights index (cited by amelius to argue the US scores higher) is methodologically sound, which is roughly where you would expect a 600-comment thread to end up.
What the Sovereignty Argument Actually Resolves To
Here is the thing both sides of that debate keep skating past: open weights make the sovereignty question answerable in a way closed APIs never can. If the file is MIT-licensed, you are not choosing between Beijing and Washington — you are choosing a host, and the host can be in your own jurisdiction, or your own rack. That is not hypothetical positioning either. On October 17, the same day V4.1-Flash landed, NVIDIA released three open Nemotron models pitched explicitly at enterprises that want agents running locally for data-security reasons, and Google made Gemini 4 Argon globally available for enterprise orchestration. The same week, Microsoft moved Copilot Studio agent accounts, shared memory, and triggers into preview. The market has decided that "where does this run" is now a product category.
Where It Disappoints
The thread is not a fan club. pmoriarty relayed a claim that V4.1 Flash "hallucinates like crazy," with no benchmark attached — note it, don't bank on it. sunaookami said it takes the longest and overthinks, which pimeys rebutted by pointing out it's fast if you don't pin thinking to maximum; that exchange is really a discovery that the 1–100 reasoning-effort dial exists and that defaults matter. Betelbuddy raised the broader caution that visible chain-of-thought may be post-hoc rationalization rather than a window into the computation. And trueno offered the best phrase in the thread for what cheap tokens do to codebases: an "additive ship of theseus," software grown rather than designed, one careless generation at a time. On the positive side, Philip-J-Fry's compliment is worth repeating because it is specific — the model "doesn't try to be too clever" and generates code "that a developer would write," which is a quality no benchmark in the model card measures.
What This Means If You Write Code for a Living
- If you pay for API tokens: the 13x-to-17x price gap against a top-tier closed model is now attached to credible agentic scores — 90.6% on Terminal-Bench 2.1. Route your bulk work here and keep the expensive model for the decisions.
- If you care about data residency: the MIT license means you can stop negotiating with a vendor and start shopping for a host. Ask any provider about zero-data-retention in writing.
- If you wanted to run it at home: wait. The Q2_K quantization alone is 264.5GB, and llama.cpp support is blocked on a model-class rename that will probably be fixed before your hardware budget is.
- If you are evaluating it: set reasoning effort deliberately. Two commenters reached opposite conclusions about the same model's speed because one of them left it maxed.
The thing to watch over the next few weeks is not the ranking — a 2.7x daily lead does not evaporate quickly — but the tooling. The moment llama.cpp accepts the V4.1 class name and Ollama ships support, the number of people who can run a frontier-class model with no network connection goes from a few hundred to a few hundred thousand, and the sovereignty debate stops being about trust and starts being about hardware budgets. October 2026 has already delivered 14 new models from 13 providers. The one that mattered showed up as a 510GB file with a date in its name.
-EditorZ
Photo by Lightsaber Collection on Unsplash
Post a Comment