Agentic work
Near frontierGDPval-AA v2 Elo
Independent AA. 0731 jumps 1189 → 1559; second-highest open-weights once released, behind Kimi K3.
SPIRITT
DeepSeek V4 FlashDeepSeek V4 Flash is cheap agent speed
The July 31, 2026 official 0731 build jumps ~10 AA Intelligence points over the April Flash preview, lands near GPT-5.6 Luna on intelligence at a fraction of the cost, and ships MIT open weights (284B total / 13B active) with a 1M context. On SPIRITT you put that efficiency inside a real agent workspace.
Across X on DeepSeek V4 Flash
Local runs, cache economics, always-use-Flash routing advice, and open-weight hosting vibes.
Official launch card
DeepSeek's official Flash release: same tiny active footprint, much stronger agent post-training, DSpark speculative decoding, Responses API / Codex support.
What DeepSeek built it for
A 284B MoE with only 13B active parameters, 1M context, MIT weights, and first-party API pricing at $0.14 / $0.28 per M tokens with an extreme ~98% cache-hit discount.
0731 is a post-training + agentic coding jump over the April preview: higher Terminal-Bench / DeepSWE / GDPval-AA, lower token use, and explicit coding-agent integrations—not a new pretrain size.
MIT open weights13B active1M context98% cache hitsAgentic codingCodex-ready
AA Intelligence Index
AA article: Flash 0731 ~50 (model page max-effort snapshots also show ~52). Near Luna/GLM/Muse 1.1, behind Kimi K3 open-weights lead.
Where V4 Flash is strong vs other models
Same score means nothing alone. These are head-to-head charts so you can see where it leads, where it is close, and what that means for real agent work.
Terminal agents
Near frontierTerminal-Bench 2.1
AA: 79% (+17 pts vs predecessor). Vendor HF table also lists higher harness figures—prefer AA for cross-lab compare.
Science coding
SciCode (AA)
AA: SciCode ~50% (+5 vs predecessor).
Hard exams
Humanity's Last Exam (AA)
AA: HLE ~37% (+5 pts).
Speed
Near frontierOutput tokens / sec (AA)
AA first-party API ~122 tok/s—fast for its intelligence band.
Price
WinnerAPI input ($ / M tokens)
$0.14 / $0.28 list with ~$0.003 cache hits (~98% discount). Core reason AA cost/task undercuts Luna.
Efficiency
WinnerAA Index output tokens (millions)
0731 used ~206M output tokens vs ~234M for prior Flash (−12%) while scoring higher.
Honesty gap
AA-Omniscience hallucination %
Hallucination falls (~96%→84%) but remains high; accuracy flat. Strong agents still need verification loops.
Sources: Artificial Analysis DeepSeek V4 Flash 0731 article and model page; DeepSeek API changelog / HF DeepSeek-V4-Flash-0731 card. AA composite ~50; some max-effort snapshots read ~52. Vendor Terminal-Bench tables can exceed AA harness numbers. On SPIRITT, pair Flash with tools and checks—not blind trust.
How It Works
From zero to DeepSeek V4 Flash running real work in a cloud agent environment
01
Open a workspace
Open a workspace and land in a fully equipped cloud computer: browser, files, terminal, integrations, and memory. No local setup. No thin chat box pretending to be an agent.
02
Pick DeepSeek V4 Flash
Open the model picker and choose DeepSeek V4 Flash. Same model shipped for agentic work, now inside a workspace that already has tools, browser control, and durable context.
03
Build or automate
Tell it what to ship or what to run. V4 Flash can code, call tools, drive the browser, coordinate multi-step work, and keep going while you step away. The point is not another chat window. It is an agentic environment where V4 Flash actually does the job.
Try DeepSeek V4 Flash in an agentic environment
Cheap, fast, open-weight agent horsepower inside a full cloud workspace.
Questions
Frequently asked questions
01What is DeepSeek V4 Flash?+
DeepSeek's efficient V4 Flash tier (official 0731 build): 284B total / 13B active MoE, 1M context, MIT open weights, and very low API prices with aggressive cache discounts. Built for high-volume agent and coding work.
02How is 0731 different from the April Flash preview?+
Same size and price, much stronger agent post-training: +10 AA Intelligence points, big GDPval-AA and Terminal-Bench jumps, fewer output tokens, Responses API / Codex support, and public MIT weights.
03What is V4 Flash best for?+
Cost-sensitive agent loops, coding assistants, batch tool use, and teams that want open weights or extreme cache economics without paying frontier list prices.
04What are some of the main limitations of this model?+
Text-only I/O on the AA card, still verbose vs medians, hallucination rates remain high even after improvement, and it trails Kimi K3 / closed frontier leaders on raw intelligence.
05What is a good example of using V4 Flash well?+
Run multi-step coding or ops agents in SPIRITT where cache hits dominate cost—repeat large contexts, verify outputs, and keep human checks on factual claims.
06Where is DeepSeek V4 Flash best to test for full agentic capabilities?+
SPIRITT Workspaces. Pick DeepSeek V4 Flash in the model picker and run real work in a fully equipped cloud environment: tools, browser, files, terminal, memory, and durable sessions. That is where the model can act as an agent, not just chat.
Buy from builders who use what they sellBuilt using