Up to date
deepseek-ai MIT out 13 months ago
685B parameters (8 of 256 experts per token), text in, text out, 160K tokens of context. 1 provider serves it from $0.27 per million input tokens.
Verified 2026-10-26 07:27 UTC Rank #227 in LLMs · score 87.4 method v1 Sources 4 Downloads 30d 39.9k · #397 of 619 tracked
Picked by rule from 27 tracked pieces: the lab's own post, the most shared, a field report, the latest.
DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention to optimize long-context efficiency while maintaining performance on par with V3.1-Terminus.
DeepSeek-V3.2 is an open-weights language model for text generation published by deepseek-ai and added to Hugging Bay Jun 29, 2026.
The 3 picks above, plus these 22. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
| # | Cheapest live row | In /1M | Out /1M | Context | Quant | Seen |
|---|---|---|---|---|---|---|
| 1 | SiliconFlow | $0.27 | $0.41 | 160K | fp8 | Oct 26 |
Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.
Served by 1 API provider; 3 quantized builds for local use.
Top Spaces running it: burtenshaw/karpathy-llm-council ♥ 292 · dropkickJesus999/deepsiteV3 ♥ 32 · kalhdrawi/omnidev ♥ 11 · Daksh0505/Youtube-Chatbot ♥ 5
Weights only, an estimate: leave room for the KV cache. Method
About 411 GB at INT4, 823 GB at FP8, 1.65 TB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.
Estimate params × bytes per param × 1.2 · how we estimate
Third-party measurements only; author-reported numbers stay in Reference, marked as such.
54.12. Each tick is a model on this site with a measured BFCL row; author-reported numbers stay in Reference. Percent (overall accuracy). FC
Source · leaderboard · 2026-10-26
Weights, every dated article, and repo commits on one rail.
Teal is the weights (2025-09-29), amber triangles are 9 dated articles (each links to it), grey ticks are repo commits.
19 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.
Everything else, folded: the full tables, every source.
| Benchmark | Value | Setting / notes | Reported by | Date |
|---|---|---|---|---|
| measured BFCL · Call tools and APIs | 54.12 percent (overall accuracy) | FC5 sub-scores
| leaderboard | 2026-10-26 |
| author-reported AIME 2025 · Reason through hard problems | 89.3 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported Aider Polyglot · Write code with agents | 74.5 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported BrowseComp | 40.1 | Agentic Tool Use | model author | 2026-10-07 |
| author-reported BrowseComp-zh | 47.9 | Agentic Tool Use | model author | 2026-10-07 |
| author-reported Codeforces · Write code with agents | 2121 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported GPQA Diamond · Reason through hard problems | 79.9 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported HMMT · Reason through hard problems | 83.6 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported Humanity's Last Exam · Reason through hard problems | 19.8 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported LiveCodeBench · Write code with agents | 74.1 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported MMLU-Pro · Reason through hard problems | 85 | Reasoning Mode w/o Tool Use | model author | 2026-10-07 |
| author-reported SWE-bench Multilingual · Write code with agents | 57.9 | Agentic Tool Use | model author | 2026-10-07 |
| author-reported SWE-bench Verified · Write code with agents | 67.8 | Agentic Tool Use | model author | 2026-10-07 |
| author-reported SimpleQA · Reason through hard problems | 97.1 | Agentic Tool Use | model author | 2026-10-07 |
| author-reported Terminal-Bench · Write code with agents | 37.7 | Agentic Tool Use | model author | 2026-10-07 |
Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.
| Pipeline tag | text-generation |
|---|---|
| Library | transformers |
| Architectures | DeepseekV32ForCausalLM |
| Model type | deepseek_v32 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 3,946,184,704 · F32: 42,667,040 · F8_E4M3: 681,408,069,632 |
| Size band | XXL |
| MoE | yes |
| Experts | 256 |
| Experts per token | 8 |
| Context length | 163,840 (config.json max_position_embeddings) |
| License | MIT (mit) |
| Hub snapshot | 2026-10-26 |
vLLM: supported · DeepseekV32ForCausalLM · 4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60SGLang: not in registry · 60d941293e3af27a12c6eb64e2c3e2e1a17f100f
vLLM registry (4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60) · SGLang registry (60d941293e3af27a12c6eb64e2c3e2e1a17f100f) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.
Weights-only memory for DeepSeek-V3.2-Exp: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 1.65 TB · Cluster (≥640 GB) |
|---|---|
| FP8 (1 byte/param) | 823 GB · Cluster (≥640 GB) |
| INT4 (0.5 bytes/param) | 411 GB · Multi-GPU node (<640 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.
Base model: deepseek-ai/DeepSeek-V3.2-Exp-Base
quantized: 3finetune: 13adapter: 3
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| QuantTrio/DeepSeek-V3.2-Exp-AWQ | quantized | AWQ | 685B | XXL | mit | QuantTrio | 4 | 10.4k |
| QuantTrio/DeepSeek-V3.2-Exp-AWQ-Lite | quantized | AWQ | 685B | XXL | mit | QuantTrio | 4 | 665 |
| Tile-AI/DeepSeek-V3.2-Exp-TileRT | finetune | — | 697B | XXL | mit | Tile-AI | 11 | 0 |
| libraos/DeepSeek-V3.2-Exp-AWQ | quantized | AWQ | 685B | XXL | mit | libraos | 0 | 113 |
| facadelighting/facadelightai | finetune | — | 4.76B | M | other | facadelighting | 0 | 55 |
| altf4isuseful/justatestok | finetune | — | — | — | openrail | altf4isuseful | 1 | 0 |
| Prodigy16/Humanforcast | finetune | — | — | — | apache-2.0 | Prodigy16 | 0 | 0 |
| majamaja818/brains | finetune | — | — | — | llama3.3 | majamaja818 | 0 | 0 |
| YOBIZ/Kael | finetune | — | — | — | unknown | YOBIZ | 0 | 0 |
| dgknknk/HBA | finetune | — | — | — | apache-2.0 | dgknknk | 0 | 0 |
| Nateturk/moneymaker1 | finetune | — | — | — | other | Nateturk | 0 | 0 |
| aqzxsw12345/qwer | adapter | — | — | — | apache-2.0 | aqzxsw12345 | 0 | 0 |
| Day2507/IA_Day | finetune | — | — | — | cc | Day2507 | 0 | 0 |
| afsdfdsfds/ChatYapEnhance | adapter | — | — | — | apache-2.0 | afsdfdsfds | 0 | 0 |
| Bman2001/Medai | finetune | — | — | — | — | Bman2001 | 0 | 0 |
| moreiraj93/moreiraj93 | finetune | — | — | — | mit | moreiraj93 | 0 | 0 |
| Mechai1230/StreetFighter1VoiceMan | finetune | — | — | — | gpl | Mechai1230 | 0 | 0 |
| fabrioop/BlaxTOP | adapter | — | — | — | mit | fabrioop | 0 | 0 |
| 35u394u50943u609234/notes_to_song | finetune | — | — | — | openrail | 35u394u50943u609234 | 0 | 0 |
| Date | Commit | SHA |
|---|---|---|
| 2025-11-18 | Update README | 194c67e |
| 2025-11-17 | fix indexer rope | 1938e9d |
| 2025-10-09 | Update README.md (#12) | 9d2f599 |
| 2025-10-09 | Update SGLang server starting command (#13) | 8a4d491 |
| 2025-09-29 | fix a typo of tilelang ref (#11) | c60171d |
| 2025-09-29 | Update README.md (#4) | 0770bf3 |
| 2025-09-29 | docs: Update README with HuggingFace and SGLang instructions (#3) | a678d82 |
| 2025-09-29 | update readme | 97efda3 |
| 2025-09-29 | update readme | 2969437 |
| 2025-09-29 | add figure | 72cacec |
| 2025-09-29 | add readme | 600c271 |
| 2025-09-29 | add inference demo | 3f13c8c |
| 2025-09-29 | Add files using upload-large-folder tool | e41a37b |
| 2025-09-29 | Add files using upload-large-folder tool | 8b7dbbd |
| 2025-09-29 | Add files using upload-large-folder tool | c22a66f |
| 2025-09-29 | Add files using upload-large-folder tool | dd0d1d5 |
| 2025-09-29 | initial commit | 8d75a26 |
Hosted endpoints observed for DeepSeek-V3.2-Exp, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.27/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| SiliconFlow openrouter | deepseek/deepseek-v3.2-exp | 160K | fp8 | $0.27 | $0.41 | — | — | yes | live | 2026-10-26 |
Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Hub snapshot: 2026-10-26 · HF API: 2026-10-26 07:18 UTC · Providers: 2026-10-26 07:26 UTC · Articles: 2026-10-26 07:27 UTC · Summary: 2026-10-07 15:04 UTC · Last verified 2026-10-26 07:27 UTC
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.
