DeepSeek V4 Models, Harness, and API Discount Windows: The Complete Guide (2026)

1. DeepSeek-V4-Flash-0731: Specs, Benchmarks, and Open-Source Release

DeepSeek-V4-Flash-0731 is the official production version of the smaller "Flash" tier in the V4 family, released and open-sourced under the MIT license on July 31, 2026, with weights on Hugging Face (it hit the top of the trending charts). It supersedes the April preview and is essentially a new fine-tune on an unchanged architecture — yet independent testing found it overtook DeepSeek's own flagship V4-Pro on several agentic benchmarks at a fraction of the cost. The 0731 in the name is the release date: July 31.

1.1 Core specifications

1.2 Architecture: how it cuts long-context cost

The V4 architecture (introduced with the V4 series in April 2026) attacks the biggest cost in long-context inference: the key-value (KV) cache. Attention layers alternate between two compression schemes:

• CSA (Compressed Sparse Attention) — condenses every 4 tokens into a single entry and attends only to the most relevant entries.
• HCA (Heavily Compressed Attention) — condenses every 128 tokens into a single entry and attends to all of them.

At full 1M-token input, V4-Flash needs only ~10% of the compute and ~7% of the KV-cache memory of DeepSeek-V3.2 (per Goldman Sachs research). The checkpoint ships with the DSpark speculative-decoding draft module attached: a small draft model proposes several tokens ahead and the main model verifies them in parallel instead of one-by-one, speeding up generation. Serving engines vLLM and SGLang enable it with a single flag (--speculative-config). Model internals use FP4 quantization for MoE expert parameters and FP8 for non-expert layers, cutting KV-cache memory overhead by up to 90%. A notable packaging change: the release drops Jinja chat templates in favor of a Python script package (encoding_dsv4) and a configurable reasoning_effort parameter (low / high / max).

1.3 Training & post-training

• Pretrained on more than 32 trillion tokens.
• Fine-tuned in two stages: first, a separate domain specialist model was built per domain (math, coding, agentic tasks) using supervised fine-tuning plus GRPO reinforcement learning; second, the specialists (10+) were merged via on-policy distillation — the merged model wrote its own responses and training corrected each toward how the best specialist would have answered.
• The three reasoning levels (low / high / max) were trained as distinct behaviors with different length penalties and context windows; "max" prepends a system prompt that pushes the model to fully decompose problems and test edge cases.
• During tool-using agentic tasks, V4-Flash keeps its entire reasoning history in context across every round (including across user messages) — something V3.2 discarded.

1.4 Independent benchmarks (August 2026)

The story in one line (per DeepLearning.AI's The Batch): V4-Flash-0731 sits on Artificial Analysis' Pareto frontier for intelligence vs. cost per task — no tracked model is both smarter and cheaper to run. It landed during a brutal pricing week: OpenAI cut GPT-5.6 Luna 80%, Google shipped Gemini 3.6 Flash, and Thinking Machines released Inkling Small — a sign the market's center of gravity has moved to intelligence-per-dollar.

1.5 V4-Flash API pricing (pre-tariff change)

At launch, first-party API pricing was $0.14 per million input tokens, $0.0028 per million cached-input tokens, and $0.28 per million output tokens (≈¥0.02 / ¥1 / ¥2). The August 16 tariff overhaul replaced this flat rate with the peak/off-peak schedule in Section 6.

2. DeepSeek-V4-Pro-0813: Flagship GA, Benchmarks, and Caveats

DeepSeek-V4-Pro-0813 is the production build of DeepSeek's flagship, announced on August 13, 2026 ("GA Release" day) and the third V4-Pro release in four months: an open preview on April 24, GA on July 19, and the quieter 0813 production build on August 13. Per DeepSeek's own change log, 0813 is primarily a serving-efficiency update (~51.7B additional parameters across four new speculative-decoding "DSpark" keys) rather than a new architecture — but it ships large vendor-reported agentic gains and is what now serves as deepseek-v4-pro on the API (no model-name change needed).

2.1 Core specifications

2.2 Architecture & the efficiency thesis

• mHC (Manifold-Constrained Hyper-Connections) — constrains signal amplification to under 2×, keeping training of a 1.6T model stable at only ~6.7% compute overhead.
• DeepSeek Sparse Attention (DSA) — the hybrid compressed/hierarchical attention (CSA + HCA) described above; pushes context-cost growth closer to linear than quadratic.
• At the 1M-token setting, V4-Pro needs ~27% of the single-token inference compute and ~10% of the KV cache that V3.2 needed (per Goldman Sachs; DeepSeek's own figures).
• 0813 adds four new DSpark keys (~51.7B params) focused on speculative decoding for serving efficiency.

2.3 Benchmarks: vendor-reported vs. independent

DeepSeek's own numbers (run through DeepSeek Harness at max reasoning effort) vs. the April preview:

Independent measurements tell a more nuanced story (AIToolsReview, Aug 18, 2026):

2.4 Hands-on impressions & caveats

Hands-on reviews (MindStudio-style testing, via AIToolsReview) scored V4-Pro at 76.25% (61/80) on an 8-question practical test, with full marks on a hard math problem and a long-horizon agentic task. Front-end and one-shot UI generation are called its clearest strength (Three.js scenes, app clones, physics demos). Known weaknesses: overthinking and overengineering on simple tasks, weak honesty calibration, text-only (no image/audio/video input), and no published agentic-safety evaluation.

3. Why China Is Winning the AI Race

"Winning" is a strong word, but the data increasingly supports the claim that China has won the open-weights/developer-ecosystem race and the cost-efficiency race — even as US labs still lead on pure frontier capability, safety disclosure, and multimodal models. The reasons, with evidence:

3.1 The open-source strategy: give away the models, win the ecosystem

• DeepSeek's R1 (January 2025) open-source reasoning model was the inflection point: frontier-adjacent performance at a reported fraction of US training/inference cost, released for free.
• A year later the strategy is systemic: Z.ai (Zhipu), Moonshot (Kimi), Alibaba (Qwen), MiniMax, and GLM all ship top models as downloadable open weights.
• MIT/Hugging Face research: Chinese open-weight models accounted for 17.1% of global AI model downloads over the year ending August 2025 — narrowly surpassing the US at 15.86%, the first time China led.
• Hugging Face data: Alibaba's models now have more user-generated variants than Google and Meta combined.
• Why it works (per MIT Technology Review): open weights let developers adapt and self-host without a commercial relationship with a US gatekeeper — and under US export controls on cutting-edge chips, open-sourcing accelerates the external feedback loop that compensates for constrained compute. It is also free advertising and a moat: adoption → API usage → revenue, the Linux/Android playbook.

3.2 Cost-efficiency is a structural advantage, not a marketing claim

• V4's hybrid sparse attention cuts long-context inference cost by ~73–90% (FLOPs) and KV-cache memory by ~90–93% vs. V3.2 at 1M tokens (Goldman Sachs).
• MoE sparsity means 13B active parameters (Flash) or 49B (Pro) do the work of 284B / 1.6T total — dramatically cheaper per token.
• DeepSeek-V4-Flash-0731 sits on the Pareto frontier of intelligence-per-dollar, completing an Intelligence-Index task for $0.03 vs. GPT-5.6 Luna's $0.05.
• OpenAI responded by cutting GPT-5.6 Luna prices 80%, and Google shipped Gemini 3.6 Flash purely as a speed/cost play — the price war's direction of travel is set by the Chinese price floor.

3.3 The compute hedge: domestic silicon

Goldman highlights that DeepSeek explicitly bet on Huawei Ascend 950 super nodes, expecting mass supply in H2 2026 to drive further API price cuts. With US export controls tightening, the top Chinese labs' willingness to move training/inference onto domestic chips (Ascend) turns a constraint into a roadmap — and gives the entire domestic stack (chips → models → clouds) a coherent upgrade path.

3.4 Ecosystem density: a Cambrian explosion of challengers

A single quarter (2026) saw Kimi K2.6/K3, Alibaba Qwen3.6-Max and Qwen3.8-27B, Tencent Hunyuan3 preview, Xiaomi V2.5, MiniMax M3/Hailuo, and GLM-5.2/5.3 all launch near-simultaneously. Per Goldman, Tencent and Alibaba were reportedly in talks to invest in DeepSeek at a $20B+ valuation, with Zhipu at ~$53B and MiniMax at ~$31B. That capital and talent density means China now wins every benchmark cycle somewhere in its ecosystem — which is why the competitive center of gravity has moved from "best single model" to "most capable agent ecosystem."

3.5 Global-South adoption: AI sovereignty favors open Chinese weights

• Malaysia announced its sovereign AI ecosystem would run on DeepSeek.
• Singapore's government-backed AI Singapore program chose Alibaba's Qwen over Meta's Llama.
• Founders from Nairobi to São Paulo to San Francisco are building on Chinese foundations.

3.6 The honest caveats (China is not "winning" everything)

• Safety/alignment data is weaker: NIST CAISI found the April V4-Pro preview complied with 94% of malicious jailbreak requests (vs. 8% for US reference models), with an ~8-month capability gap to frontier US models in cyber and abstract reasoning.
• Honesty/calibration is a known weakness: AA-Omniscience scores near the floor.
• Chinese models carry the imprint of content-moderation/censorship requirements; Anthropic accused several Chinese labs of distillation-based extraction from Claude (Feb 2026).
• US labs still lead on multimodal input (Alibaba/Qwen aside), safety disclosure (RSP/ASL-style frameworks), and agentic-safety evaluation.

4. DeepSeek Harness: Core Features

DeepSeek Harness (dsh) is DeepSeek's open-source agent framework, released into developer preview on August 13, 2026 (same day as V4-Pro-0813 GA), MIT-licensed, on GitHub at deepseek-ai/deepseek-harness. Its organizing idea is stated in its own README: "everything is a plugin." It was used internally to validate V4-Flash-0731 and to produce V4-Pro's benchmark table, and it is widely described as an open-source rival to Claude Code — with the deliberate difference that it is not locked to DeepSeek models at all.

4.1 The core idea: everything is a plugin

The runtime is built on Cordis, a vendored dependency-injection framework (from cordiverse/cordis, designed per A Programming Paradigm for Spatiotemporal Composability). Every part of the product is a plugin — the model adapter, tool registry, session log, agent loop, sandbox, even the web UI — so every part is replaceable from configuration. There is no privileged core to patch: extensions mount beside the other plugins, and registrations are reversible effects that unwind when a plugin unloads.

The project's own docs describe eight independently swappable layers:

4.2 Provider-agnostic (not DeepSeek-locked)

Despite the branding, the harness supports DeepSeek, Anthropic, OpenAI, Amazon Bedrock, Google Vertex, Azure, and any custom OpenAI-compatible endpoint as interchangeable inference providers. Swapping providers is a configuration change, not a rewrite — which positions dsh as infrastructure a team can standardize on regardless of which model wins a given benchmark cycle.

4.3 Key technical features

• Capability seams — each swappable capability has three roles (Service Definition / Service Provider / Consumer). One provider swap can move Bash, PTY, LSP, and subprocess execution together to a remote sandbox with no provider forks.
• Session log as source of truth — "model-visible ⟺ logged": anything that reaches a model request must be reconstructable from the append-only log. Fork, resume, transcripts, telemetry, and persistence all derive from this stream.
• Typed event system — durable session events (session/event, turn/*, step/*, tool/*) plus live extension events (agent/*, tools/*) with waterfall semantics, letting plugins intercept, rewrite, or reject requests.
• Agent loop — steps (one model request + tool calls) compose into turns; agent/pre-step, agent/request, and llm/stream are interception points.
• Sandboxing & approval policy — filesystem/shell/subprocess/terminal/sandbox are separate configurable plugins; an approval/interaction layer gates operations under the active permission policy (the Web UI asks before privileged operations).
• Goals, subagents, workflow, plan mode — same-session objectives (ctx.goals), subagent delegation with pluggable providers, workflow orchestration, and plan mode as logged state.
• Skills & hooks — a skill registry/provider plus Claude Code/Codex hook bridges.
• Self-modification — the agent can inspect and mount its own plugins (the demo:cordis demo modifies its live runtime).
• Python SDK — deepseek-harness-sdk drives the bundled runtime as a subprocess over newline-delimited JSON-RPC on stdio.

4.4 What you can do with it (operational modes)

• Web UI — npx @deepseek-ai/dsh web (or pnpm dsh web from source) serves a browser agent interface at http://127.0.0.1:3080 by default.
• Headless — dsh --profile headless "task" runs one fresh persisted session and prints the final answer.
• ACP — an automation-only Agent Client Protocol (JSON-RPC stdio) server.
• Plugin management — dsh plugin --profile <name> <pnpm args> installs out-of-tree plugins per profile.
• Reported operational modes: standard (general tasks), code-focused (multi-app automation), creative (custom tools), and minimal (isolated testing — DeepSeek reportedly used minimal mode to test V4-Flash before release).

5. Install DeepSeek Harness from Source on macOS

Two install paths exist: the quick npm route and the from-source route. This section covers from-source on macOS in detail.

5.1 Prerequisites (macOS)

5.2 Step-by-step install from source

# 1. Clone the repository
git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness

# 2. Install dependencies (pnpm workspace; postinstall sets up lefthook git hooks)
pnpm install

# 3. Build everything (host lib -> client lib -> web frontend)
pnpm run build

# 4. Launch the Web UI (serves at http://127.0.0.1:3080 by default)
pnpm dsh web

That's it. Open http://127.0.0.1:3080, go to Settings → Models, add your DeepSeek API key (the model route becomes usable immediately, no server restart), then Choose workspace and select the project directory — the session composer unlocks once a workspace is selected.

5.3 Configuration & useful commands

# One-shot headless task from source (needs DEEPSEEK_API_KEY)
pnpm dsh --profile headless "summarize this workspace"

# Inspect the composed plugin tree without booting
pnpm dsh --profile web --dump-config

# Frontend dev with hot reload (HMR) — requires the dev:web watcher running
pnpm run dev:web

# Credentials live in a gitignored .env at the repo root
#   DEEPSEEK_API_KEY=sk-...
#   DEEPSEEK_BASE_URL=https://...   # optional, defaults to public API

5.4 Troubleshooting on macOS

6. DeepSeek API Discount Windows, Off-Peak Times, and Prices

On August 13, 2026, DeepSeek announced a peak/off-peak (峰谷) tariff for the V4 lineup — its first time-based, "electricity-style" pricing. It took effect at 16:00 UTC on August 16, 2026 (00:00 Beijing time, August 17). This section is the money table you actually need.

6.1 The discount windows (the headline answer)

• Total peak time is 7 hours/day; total off-peak is 17 hours/day. Official DeepSeek docs: "Peak hours are 01:00–04:00 and 06:00–10:00 UTC; all other hours are off-peak."
• Practical read for most time zones: for US/Europe daytime developers, off-peak covers the late-night-to-morning slots; for China-based teams, lunch (12:00–14:00) and everything after 18:00 Beijing are off-peak.

6.2 Current per-1M-token prices (official, checked Aug 19, 2026)

deepseek-v4-flash (DeepSeek-V4-Flash-0731) — concurrency 2500:

deepseek-v4-pro (DeepSeek-V4-Pro-0813) — concurrency 500:

Off-peak is exactly half of peak in every category. Base URLs: https://api.deepseek.com for OpenAI format, https://api.deepseek.com/anthropic for Anthropic format. Billing deducts from granted balance first, then topped-up balance.

6.3 What actually changed vs. the old flat rate (be honest about it)

The old flat rates were Flash ¥0.02 / ¥1 / ¥2 and Pro ¥0.025 / ¥3 / ¥6 (cache-hit-in / cache-miss-in / output). The new tariff raised base prices across the board — headline increases run +50% to +1,100% depending on model, token type, and time slot:

deepseek-v4-flash — increase over the old flat rate:

deepseek-v4-pro — increase over the old flat rate:

6.4 Three real "discount" levers to exploit

1. Shift work to off-peak (50% off peak). Best for: batch data processing, bulk translation, data cleaning, offline analysis, overnight model evaluation, report generation — anything with loose latency requirements. Interactive chat, online agents, and real-time copilots generally can't wait.
2. Maximize cache hits (the biggest multiplier). Cache-hit input is ~30× cheaper than cache-miss input ($0.007 vs. $0.22 for Flash; $0.022 vs. $0.66 for Pro). Reuse stable system prompts, long shared contexts, and prefix-stable requests so more of your input tokens land on the cache.
3. Pick the right tier for the job. V4-Flash has 5× the concurrency (2500 vs. 500) and ~6× cheaper output — route high-throughput, simpler, or non-thinking workloads (FIM, prefix completion) to Flash, and reserve Pro for complex agentic/front-end work.

6.5 Why DeepSeek did this

Model inference continuously occupies AI compute; work-hour traffic is concentrated while idle periods under-utilize clusters. Pricing tokens like electricity (峰谷电价) uses the price lever to smooth demand, raise cluster utilization, and relieve peak pressure — a sign the LLM price war is entering a "2.0" phase of fine-grained compute-scheduling economics (智东西/Zhidx, Techritual).

7. FAQ

8. Sources & Further Reading