DeepSeek-V4-Flash
One-line: Compact 13B-active MoE variant of the DeepSeek-V4 line — hits 79.0 SWE-bench Verified (Flash-Max) and 91.6 LiveCodeBench while only activating 13B per token; weights are FP4+FP8 mixed and require custom inference.
Coding benchmarks
| Benchmark | Score | Source |
|---|---|---|
| HumanEval | — | not published for this variant |
| HumanEval+ | — | not published |
| MBPP | — | not published for this variant |
| LiveCodeBench | 91.6 | lab (V4-Flash Max mode) |
| Aider polyglot | — | not published |
| SWE-bench Verified | 79.0 | lab (V4-Flash Max mode) |
| SWE-bench Multilingual | 73.3 | lab (V4-Flash Max mode) |
| Codeforces (rating) | 3052 | lab (V4-Flash Max mode) |
LiveCodeBench rises sharply across reasoning modes: Non-Think 55.2 → Think High 88.4 → Think Max 91.6 — at High mode Flash already approaches Opus-4.6 Max (88.8) and Gemini-3.1-Pro High (91.7).
Agentic / tool-use benchmarks
| Benchmark | Score | Source |
|---|---|---|
| BFCL | — | not published for this variant |
| τ-Bench | — | not published |
| ToolACE | — | not published |
| GAIA | — | not published |
| Terminal Bench 2.0 | 56.9 | lab (V4-Flash Max mode) |
| SWE Pro | 52.6 | lab (V4-Flash Max mode) |
| BrowseComp | 73.2 | lab (V4-Flash Max mode) |
| MCPAtlas (Public) | 69.0 | lab (V4-Flash Max mode) |
| Toolathlon | 47.8 | lab (V4-Flash Max mode) |
Agentic scores are competitive with much larger open-weight peers; SWE Pro at 52.6 is strong for a 13B-active model. BFCL not published separately — when running locally, tool-calling capability should be smoke-tested against your own function schemas before deploying.
Reasoning (context only)
| Benchmark | Score | Source |
|---|---|---|
| MMLU-Pro | 86.2 | lab (V4-Flash Max mode) |
| GPQA Diamond | 88.1 | lab (V4-Flash Max mode) |
| MATH | — | not published for this variant |
| AIME 2025 | — | not published; HMMT 2026 Feb = 94.8 |
| HLE | 34.8 | lab (V4-Flash Max mode) |
Efficiency
- Active params: 13B (per token — top-6 of 256 routed + 1 shared)
- Total params: 284B (full disk-resident weight footprint)
- Context length: 1,048,576 tokens (1M) via YaRN rope scaling; recommend ≥384K context for Think Max mode
- VRAM (fp16 rough): ~26 GB active-params equivalent at runtime; full weights ~284 GB on disk (FP4 experts + FP8 other)
- Precision: MoE expert weights are FP4, most other parameters are FP8 — requires DeepSeek’s custom
encoding+inferencescripts (vLLM/SGLang/TensorRT-LLM support is preliminary at release time) - Architecture: Hybrid attention — Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA); Manifold-Constrained Hyper-Connections (mHC); Muon optimizer in pretraining
What makes it notable
DeepSeek-V4-Flash is the smaller of two MoE siblings in the V4 line and is the first open-weight release where the Flash-Max mode pushes SWE-bench Verified to 79.0 at only 13B active params per token — within ~2 points of Opus-4.6 Max (80.8) and ahead of GPT-5.4 (not published), Gemini-3.1-Pro High (80.6), and K2.6 Thinking (80.2). LiveCodeBench at 91.6 actually beats all closed-source peers in the table, including Gemini-3.1-Pro High (91.7 — virtually tied). The 1M-token context with hybrid CSA+HCA attention keeps long-context inference cheap: 27% of single-token FLOPs and 10% of KV cache vs DeepSeek-V3.2. Practical caveat — full weights are ~284 GB on disk and the FP4-expert quantization means you cannot use stock vLLM yet; budget for the official inference/ scripts and a multi-GPU host (8× H100 80GB minimum for comfortable inference, 4× H100 only at low concurrency).