DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

TL;DR

DeepSeek-V3.2 is the production successor to V3.1-Terminus built around three changes that the paper frames as the open-source gap-closers vs. GPT-5 / Gemini-3-Pro / Claude-4.5: (1) DeepSeek Sparse Attention (DSA) — a lightning-indexer + top-k token-selection mechanism instantiated under MLA that drops core-attention cost from O(L²) to O(Lk) with no measurable quality loss at 128K context; (2) a scaled GRPO recipe (post-training compute >10% of pretraining) with four stability mechanisms — unbiased K3 KL estimator, off-policy sequence masking on negative-advantage trajectories, Keep Routing for MoE expert paths, and Keep Sampling Mask for top-p/top-k truncation parity between sampler and trainer; (3) a large-scale agentic task synthesis pipeline — 1,827 synthesized environments + 85,000 prompts across code (24,667 tasks), search (50,275), general (4,417), and code-interpreter (5,908) agents, with a multi-agent verifier loop for search-agent data. A high-compute variant DeepSeek-V3.2-Speciale (relaxed length penalty during RL + DeepSeekMath-V2 proof data) reports gold-medal performance at IMO 2025, IOI 2025, ICPC World Final 2025, and CMO 2025, with Codeforces rating 2708 — on par with Gemini-3.0-Pro.

Key claims

DeepSeek Sparse Attention computes a per-query lightning indexer score I_{t,s} = Σ_j w^I_{t,j} · ReLU(q^I_{t,j} · k^I_s) (with H_I heads, ReLU for throughput, FP8 implementation) and applies attention only to the top-k key-value entries selected by I_{t,:}, instantiated under MLA’s MQA mode so the latent vector is shared across all query heads of a token [§2.1, Eq. 1–2, Fig. 2].
Continued-pretraining adapts V3.1-Terminus to DSA in two stages: a 1,000-step Dense Warm-up (only the indexer is trained; KL-divergence loss aligns indexer output to the aggregated main-attention distribution; lr 1e-3; 2.1B tokens total) followed by a Sparse Training stage (15,000 steps; indexer + main model jointly optimized but with separated computational graphs; 2048 selected KV tokens per query; lr 7.3e-6; 943.7B tokens total) [§2.1.1, Eq. 3–4].
DSA drops core-attention complexity from O(L²) to O(Lk); the lightning indexer remains O(L²) but with far fewer heads and FP8 arithmetic. End-to-end USD-per-million-tokens on H800 ($2/GPU-hour rental) shows the cost gap with V3.1-Terminus widens with context length on both prefill and decode, with the decode-side gap reaching multiples by 128K tokens [§2.3, Fig. 3].
Parity vs V3.1-Terminus is reported on three axes: matched standard benchmarks, matched ChatbotArena Elo, and same-or-better long-context evals (AA-LCR +4 pts in reasoning mode; Fiction.liveBench consistently higher) — i.e. DSA buys efficiency without measurable quality loss in the V3.2-Exp regime [§2.2].
The post-training pipeline allocates >10% of pretraining compute to RL — explicitly framed as one of the three deficiencies of prior open models (resource allocation), alongside the architecture and agent-data deficits [§1, §3].
The GRPO scaling recipe adds four named stabilizers to vanilla GRPO: (a) Unbiased KL Estimate that importance-weights Schulman’s K3 estimator to correct bias when π_θ ≪ π_ref, eliminating noisy unbounded gradients; (b) Off-Policy Sequence Masking that drops only negative-advantage sequences whose KL(π_old∥π_θ) exceeds a threshold δ, on the principle that highly off-policy negatives mislead more than they teach; (c) Keep Routing, which preserves the MoE expert paths chosen by the inference framework into the training framework so the optimized expert parameters match the sampled ones — described as crucial for MoE RL stability since DeepSeek-V3-0324; (d) Keep Sampling Mask, which propagates the top-p / top-k truncation mask from sampling into training to keep action subspaces of π_old and π_θ aligned [§3.1, Eq. 5–9].
Mixed RL merges reasoning, agent, and human-alignment into a single GRPO stage rather than multi-stage RL, explicitly to avoid catastrophic forgetting; rule-based outcome reward + length penalty + language-consistency reward on reasoning/agent tasks, generative reward model with per-prompt rubrics on general tasks [§3, Mixed RL Training].
Specialist distillation is used across six domains (math / programming / general logical reasoning / general agentic / agentic coding / agentic search) plus writing and QA; each specialist is RL-trained from the V3.2 base, then used to generate domain data for the final checkpoint, with the gap to specialists “effectively eliminated” by the subsequent RL stage [§3, Specialist Distillation].
The Thinking-Context-Management rule for tool calling departs from R1: reasoning content is discarded only when a new user message arrives; if only tool messages (tool outputs) are appended, reasoning is retained; tool-call history is preserved even when reasoning is dropped. Specifically motivated by token-waste in R1’s discard-every-turn behavior [§3.2.1, Fig. 4].
The Large-Scale Agentic Task Synthesis Pipeline has four task families with explicit (count, env-type, prompt-source) shape (Table 1): code agent (24,667, real, extracted from GitHub issue-PR pairs), search agent (50,275, real APIs, synthesized prompts), general agent (4,417, synthesized envs, synthesized prompts), code interpreter (5,908, real, extracted) [§3.2.3, Table 1].
The Search Agent data pipeline is a multi-agent loop: a question-construction agent explores long-tail entities sampled from web corpora with configurable depth/breadth, multiple heterogeneous answer-generation agents produce candidate responses, and a verification agent with search capabilities validates via multiple passes — retaining only samples where the ground-truth answer is correct and all candidates are verifiably incorrect; final reward is a generative reward model scored against multi-dimensional evaluation rubrics [§3.2.3, Search Agent].
The Code Agent environment-builder is itself an LLM agent: it mines millions of issue-PR pairs from GitHub, filters by heuristics + LLM judgment, then uses a DeepSeek-V3.2-powered environment-setup agent for package install / dependency resolution / test execution, with environments validated as “F2P>0 ∧ P2F=0” after applying the gold patch. Spans Python, Java, JavaScript, TypeScript, C, C++, Go, PHP [§3.2.3, Code Agent].
The General Agent pipeline auto-synthesizes 1,827 task-oriented environments from a “task category + sandbox-equipped” seed via an automatic environment-synthesis agent — the headline number behind the abstract’s “large-scale” claim is the count of synthesized environments, not just prompts [§3.2.3, General Agent].
Headline benchmarks on the cover figure: AIME 2025 96.0 (Speciale 96.0 / Thinking 93.1, vs GPT-5-High 94.6 / Claude-4.5 87.0 / Gemini-3-Pro 95.0); HMMT 2025 99.2 / 90.2 (vs 88.3 / 79.2 / 97.5); HLE 30.6 / 25.1 (vs 26.3 / 13.7 / 37.7); Codeforces 2708 / 2386 / 2537 / 1480 / 2701 (Speciale matches Gemini-3-Pro); SWE-Verified 73.1 / 74.9 (vs 77.2 / 76.2 / 46.4); Terminal Bench 2.0 35.2 / 42.8 (vs 54.2 / 80.3 / 80.2); τ²-Bench 84.7 / 85.4 (vs 35.2 / 29.0 / 38.6 / 36.4); Tool Decathlon strong [Fig. 1].

Method

The architecture is V3.1-Terminus with one change: dense MLA is replaced by DeepSeek Sparse Attention (DSA) instantiated under MLA’s MQA mode. Each query token computes index scores against all preceding tokens via a small lightning indexer with H_I heads, ReLU activation, and FP8 weights — chosen for throughput rather than expressivity. The top-k indices select which latent KV entries the main attention head reads. Because the latent vector is shared across all query heads of a token (MQA mode), the kernel-level requirement that each KV entry be shared across multiple queries is satisfied. The model is initialized from V3.1-Terminus at 128K context and trained in two stages: a 1,000-step dense warm-up where only the indexer is trained against a KL-divergence-to-main-attention objective (everything else frozen), then a 15,000-step sparse stage where the top-k selector is enabled (k=2048), the indexer continues training against KL-on-selected-tokens, and the main model is trained against language-modeling loss — with the indexer’s input detached from the computational graph so the two losses optimize independent paths.

Post-training is a specialist-distillation + single-stage mixed GRPO pipeline. Per domain, an RL-trained specialist (writing, QA, math, coding, general reasoning, general agentic, agentic coding, agentic search — each with thinking and non-thinking variants) generates training data for the final checkpoint, which is then RL-trained itself on a merged reasoning + agent + alignment objective. The GRPO modifications described in §3.1 — unbiased K3 KL, off-policy negative-sequence masking, Keep Routing, Keep Sampling Mask — are presented as load-bearing for scaling RL compute past 10% of pretraining. For thinking-in-tool-use, the agent-loop context-management rule retains reasoning across tool-only message turns and drops it only when a new user message arrives. Agent training data comes from the synthesis pipeline: code (real GitHub-issue environments, LLM-built sandboxes, F2P/P2F validation), search (long-tail-entity exploration + multi-agent verifier loop), general (1,827 LLM-synthesized environments and toolsets), and code-interpreter (Jupyter as the action space, real curated problems).

DeepSeek-V3.2-Speciale shares the same architecture and base but is RL’d only on reasoning data with a reduced length penalty, plus DeepSeekMath-V2’s dataset and reward method for mathematical proofs. The high-compute variant is what produces the IMO/IOI/ICPC-World-Final gold-medal numbers.

Results

Parity in standard mode: V3.2-Exp (the pre-Speciale checkpoint) matches V3.1-Terminus on diverse standard benchmarks and on ChatbotArena Elo, with no observed degradation despite the architectural change to sparse attention [§2.2].
Long-context gain in standard mode: V3.2-Exp scores +4 points over V3.1-Terminus on the independently-run AA-LCR reasoning-mode benchmark, and outperforms it across Fiction.liveBench metrics — i.e. the sparse architecture is at minimum neutral and sometimes positive on long-context tasks where it should have been most at risk [§2.2].
Inference cost on H800 ($2/GPU-h): Cost-per-million-tokens curves show DSA savings growing with position; the decode-side gap reaches multiples by 128K tokens [Fig. 3].
Final benchmarks (DeepSeek-V3.2 / Speciale variant): AIME 2025 96.0; HMMT 2025 99.2; HLE 30.6; Codeforces 2708 (rating); SWE-Verified 73.1; Terminal Bench 2.0 35.2; τ²-Bench 84.7; Tool Decathlon competitive with Gemini-3-Pro and ahead of Claude-4.5-Sonnet on multiple tool-use axes [Fig. 1].
Olympiad-level claim: Speciale reports gold-medal performance at IOI 2025, ICPC World Final 2025, IMO 2025, CMO 2025 [§1].
Agentic synthesis scale: 1,827 synthesized general-agent environments, 85,000 complex prompts across all four task families; reasoning unification is described as occurring during the cold-start phase before the RL stage [§1, Table 1].

Why it’s interesting

The interesting part isn’t a single new technique — it’s the integration. Three threads that have been developing independently across the wiki get fused into one open release:

DSA is the architecture-side answer that complements the sparse-attention thread already tracked under The Diffusion Duality-adjacent attention designs and reverse-engineering studies like Sparsity is Cool: Reverse-Engineering MoBA and NSA; the lightning-indexer + top-k formulation is the production-tuned cousin of MoBA/NSA, instantiated on MLA’s MQA mode rather than vanilla MHA. It’s also a strong external datapoint for the LLM Inference Efficiency thread, where production wall-clock has so far been about KV-cache compression and speculative decoding rather than attention-pattern sparsity.
The GRPO scaling tricks (unbiased K3 KL, off-policy negative masking, Keep Routing, Keep Sampling Mask) are the stability-side answer that the Reasoning RL concept page has been circling — concrete, named recipes for taking RL past 10% of pretraining compute without the well-known divergences. Keep Routing in particular is a direct response to the MoE-routing-stability issues this wiki has tracked under MoE Routing Design; it forces inference-time and training-time routing to agree, which is a stronger constraint than load-balance losses alone.
The agentic synthesis pipeline (1,827 environments, multi-agent search-data loop with verifier-only-retain gating, LLM-built code sandboxes) is the data-side answer that maps directly onto the Synthetic Training Data / Tool-Use Agents axis — the same “multi-agent role-play + gated reward” recipe in Close the Loop: Synthesizing Infinite Tool-Use Data via Multi-Agent Role-Playing and MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling, at production scale with broader environment coverage. This is what Dejia called “impressive” in the Slack note: the gating logic (“retain only samples where ground-truth correct AND all candidates verifiably incorrect”) is sharper than most open synth recipes, and it spans code/search/general/code-interpreter rather than one domain.

Compared to the closest open siblings in the same release window — DeepSeek-V4 collection release (Flash + Pro, up to 1.6T), GLM-4.5: Agentic Foundation Model from Z.ai, GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models, Kimi K2: Open Agentic Intelligence, Kimi K2.6: Advancing Open-Source Coding — V3.2’s contribution is the documented integration: not just an agent-model claim, but a full disclosure of the four GRPO stabilizers, the four agent-task synthesis pipelines with task counts, and an architecture that makes the long-context RL traces affordable to roll out.