DeepSeek V4 Flash 0731
MoE workstation284B total, 13B active per token (MoE). Same FP4+FP8 hybrid-attention family as V4 Pro: Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA) across 61 layers, manifold-constrained Hyper-Connections (mHC), Muon optimizer, 32T+ pretraining tokens.
- Context: 1M native, 384K max output; three modes (non-think / think-high / think-max).
The 2026-07-31 iterative update of V4 Flash (supersedes the April model). Per Artificial Analysis, a 10-point Intelligence Index jump to 50 - 6 points above V4 Pro, 1 behind GLM 5.2 / GPT-5.6 Luna, 7 behind Kimi K3. Agentic Elo 1559 (up from 1189), Terminal-Bench 2.1 79% (+17), Humanity’s Last Exam 37% (+5), GPQA-Diamond 91% (+1), SciCode 50% (+5). Token usage -12%; hallucination rate 84% (a 12-point drop); AA-Omniscience Index -16 (+7). Pricing unchanged at $0.14/$0.28 per 1M in/out (cache-hit $0.0028/M, a 98% discount). The model the deepseek-chat/reasoner API aliases now route to (retired 2026-07-24).
-
Local run (2026-07-31): Unsloth’s Dynamic 2.0 GGUFs landed local inference. UD-Q4_K_XL is a 155GB lossless 4-bit build (~168GB RAM); UD-Q8_K_XL is a 162GB 8-bit full-precision build (~175GB RAM, only 7GB bigger than Q4 because the 13B active experts dominate). Both fit two stacked DGX Sparks (256GB unified via ConnectX-7) or a 192GB+ unified rig; a smaller 3-bit (~110GB RAM) that would fit a single 128GB Spark is announced but not yet published. Run via Unsloth or
llama.cpp -hf- Ollama only ships the:cloudendpoint, so there is no local Ollama tag.
Open weights under MIT (full weights expected in the coming weeks per DeepSeek; the Unsloth GGUFs are available now).
Agents on Rails benchmark (Aug 2026, Le Mans round). 65.1% accuracy on 63 runs - last of the 16-model field, with only 12.7% API recall. The low recall suggests it often hand-rolled fixes instead of using the relevant Rails API. Still the cheapest per-token model in the field, so the gap is cost-vs-accuracy rather than capability at any price.
AI-generated content marks
This model embeds text watermarks in generated text and adds C2PA provenance metadata to supported files such as .png, .jpg, and .svg. Marks can be lost through editing, screenshots, or format conversion, so their absence does not prove a file is human-made.
- 284.0B
- 1000k
- mit
- 🇨🇳 China
- Jul 2026
What people are building with DeepSeek V4 Flash 0731
Real demos from X
Ben Davis's 2x NVIDIA DGX Spark rack hits 50+ tok/s on DeepSeek V4 Flash 0731 with multiple sessions View on X →
Benchmark scores
Vendor-reported - from the developer's own model card / tech report
Vendor-reported - from the developer's own model card / tech report
Ran this model on your own hardware? Join free and add your measured tok/s to the community numbers.
Score per dollar
3000 pts per $/M input
general_score (90) divided by cheapest input price ($0.03/M). Higher is better value. See live pricing.
Related models
Guides covering DeepSeek V4 Flash 0731
Save your hardware and every model page answers the real question: will it run on your machine, and how fast?
Join free - save your rig →Run it locally
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig | UD-Q4_K_XL | UD-Q8_K_XL |
|---|---|---|
| NVIDIA Jetson Orin NX 16GB | no -> cloud | no -> cloud |
| Jetson AGX Orin 64GB | no -> cloud | no -> cloud |
| Single GTX 1080 Ti (11GB) | no -> cloud | no -> cloud |
| Single RTX 4090 (24GB) | no -> cloud | no -> cloud |
| Single RTX 5090 (32GB) | no -> cloud | no -> cloud |
| RTX PRO 6000 Blackwell (96GB) | no -> cloud | no -> cloud |
| 2x RTX 3090 (48GB) | no -> cloud | no -> cloud |
| 2x RTX 5090 (64GB) | no -> cloud | no -> cloud |
| 4x RTX 4090 (96GB) | no -> cloud | no -> cloud |
| 4x RTX 5090 (128GB) | offload | offload |
| MacBook Pro M5 Max 128GB | no -> cloud | no -> cloud |
| Ryzen AI Max+ 395 128GB | no -> cloud | no -> cloud |
| DGX Spark 128GB unified | no -> cloud | no -> cloud |
| 4x H100 80GB (320GB) | fast 2454.2t/s | fast 993.1t/s |
| NVIDIA DGX Station 748GB | fast 1465.2t/s | fast 592.9t/s |
| 8x RTX 3090 rack (192GB) | fast 1371.7t/s | fast 555.1t/s |
| AMD Instinct MI300X (192GB) | fast 412.4t/s | fast 394.6t/s |
| Mac Studio M4 Ultra 192GB | fast 92.3t/s | fast 88.3t/s |
| Mac Studio M4 Ultra 512GB | fast 92.3t/s | fast 88.3t/s |
| Dual EPYC 9004 + 768GB DDR5-4800 | fast 35.7t/s | fast 34.2t/s |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | ok 15.9t/s | ok 15.2t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | ok 11.9t/s | ok 11.4t/s |
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
Download options
Or run it in the cloud
Live per-provider pricing, throughput and uptime - refreshed about 19 hours ago via OpenRouter. Click a column to sort.
some pricing may be stale - last verified 2026-09-17
| Provider | Type | Input $/M | Output $/M | Cache $/M | Tok/s | Latency | Uptime | Value |
|---|---|---|---|---|---|---|---|---|
|
OpenInference
|
API | 0.03 | 0.13 | 0.010 | - | - | 100.00% | best uptime |
|
DigitalOcean
stale
|
API | 0.08 | 0.25 | 0.025 | - | - | - | |
|
Nous Portal
stale
|
API | 0.11 | 0.22 | - | - | - | - | |
|
BaseTen
|
API | 0.13 | 0.26 | 0.028 | - | - | 100.00% | |
|
CoreWeave
|
API | 0.13 | 0.28 | 0.070 | - | - | 100.00% | |
|
Io Net
stale
|
API | 0.22 | 0.49 | 0.114 | - | - | 100.00% | |
|
Alibaba
|
API | 0.35 | 1.06 | 0.035 | - | - | 100.00% | |
|
Novita
|
API | 0.41 | 1.23 | 0.026 | - | - | 100.00% | |
|
Cloudflare
|
API | 0.44 | 1.32 | 0.014 | - | - | 100.00% | |
|
DeepSeek
|
API | 0.44 | 1.32 | 0.014 | - | - | 100.00% | |
|
AtlasCloud
|
API | 0.44 | 1.32 | 0.028 | - | - | 100.00% | |
|
GMICloud
|
API | 0.29 | 0.86 | 0.009 | - | - | 99.99% | |
|
Baidu
|
API | 0.44 | 1.32 | 0.014 | - | - | 99.98% | |
|
Wafer
|
API | 0.10 | 0.25 | 0.050 | - | - | 99.95% | |
|
Makora
|
API | 0.09 | 0.20 | 0.020 | - | - | 99.93% | |
|
Together
|
API | 0.14 | 0.28 | 0.030 | - | - | 99.93% | |
|
DeepInfra
|
API | 0.06 | 0.18 | 0.015 | - | - | 99.92% | |
|
Reka
|
API | 0.11 | 0.66 | 0.007 | - | - | 99.92% | |
|
DigitalOcean
|
API | 0.12 | 0.24 | 0.024 | - | - | 99.92% | |
|
Venice
|
API | 0.18 | 0.35 | 0.035 | - | - | 99.87% | |
|
Relace
|
API | 0.06 | 0.12 | 0.012 | - | - | 99.82% | |
|
SiliconFlow
|
API | 0.22 | 0.66 | 0.028 | - | - | 99.77% | |
|
NextBit
|
API | 0.35 | 1.06 | 0.012 | - | - | 99.77% | |
|
Morph
|
API | 0.14 | 0.40 | 0.036 | - | - | 99.69% | |
|
Inceptron
|
API | 0.06 | 0.20 | 0.010 | - | - | 99.68% | |
|
Phala
|
API | 0.44 | 1.32 | 0.028 | - | - | 99.54% | |
|
Fireworks
|
API | 0.22 | 0.66 | 0.007 | - | - | 99.36% | |
|
Sail Research
|
API | 0.07 | 0.34 | 0.023 | - | - | 99.22% | |
|
StreamLake
|
API | 0.06 | 0.18 | 0.002 | - | - | 99.19% | |
|
Ionstream
stale
|
API | 0.21 | 0.42 | 0.100 | - | - | 98.56% | |
|
Parasail
|
API | 0.14 | 0.28 | 0.050 | - | - | 98.17% | |
|
AkashML
|
API | 0.06 | 0.18 | 0.016 | - | - | 97.35% | |
|
Ambient
risky
|
API | 0.08 | 0.18 | 0.016 | - | - | 94.57% | |
|
Nebius
risky
stale
|
API | 0.14 | 0.28 | 0.016 | - | - | 91.05% | |
|
Decart
avoid
stale
|
API | 0.06 | 0.13 | 0.013 | - | - | 89.81% | |
|
Mancer 2
avoid
|
API | 0.20 | 0.60 | 0.012 | - | - | 84.82% |
Default order: throughput among 95%+ uptime providers, then latency; subscriptions last. Sort by any column. Subscription rows show $/mo in the Value column - per-token columns are "-". Affiliate links are marked sponsored / nofollow. Confirm current pricing on the provider's site before committing.
Detailed API pricing page + JSON endpoint →
See who runs DeepSeek in production →
Inference cost over time
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.