Qwen vs DeepSeek — which actually wins in 2026?
China's two biggest model families, and they diverge sharply. Qwen 3.7 Max leads reasoning and knowledge — GPQA 92.4 vs 90.1, MMLU-Pro 89.6 vs 87.5 — but it's closed-weight and ~6× dearer on input. DeepSeek V4 Pro is MIT open-weight, self-hostable, and far cheaper, and it ties on agentic coding (SWE-bench 80.6 vs 80.4). Choose Qwen for peak reasoning or a fixed-fee coding plan; choose DeepSeek for cost, self-hosting, and coding parity.
| Category | Winner | Margin |
|---|---|---|
| General reasoning · GPQA Diamond | AQwen | Qwen 92.4 vs DeepSeek 90.1 — a real but modest 2.3-point edge (llm-stats) |
| Broad knowledge · MMLU-Pro | AQwen | Qwen 89.6 vs DeepSeek 87.5 — same direction, same modest size |
| Agentic coding · SWE-bench Verified | ·Tie | 80.4 vs 80.6 — a 0.2-point gap is noise, not a lead |
| API cost · per 1M tokens | BDeepSeek | DeepSeek is ~6× cheaper on input and ~10× cheaper on output |
| Real cost on prose · tokenizer-adjusted | BDeepSeek | The tokenizers are near-identical on English prose (5.23 vs 5.26 chars/token), so the ~6× rate gap carries through unchanged |
| Open weights · self-hosting | BDeepSeek | DeepSeek V4 Pro ships MIT weights; Qwen3.7 Max is closed, API-only |
| Cache economics · repeat context | BDeepSeek | DeepSeek publishes a $0.003625 cache-hit rate; Alibaba doesn't itemize its cache price |
| Context window · API model | ·Tie | 1M tokens on both — no advantage either way |
| Model lineup · breadth | AQwen | Qwen sells Max, Plus, Flash, Coder and VL tiers; DeepSeek ships Pro and Flash |
| Fixed-fee coding plan · subscription | AQwen | Alibaba's Coding Plan Pro is ¥200/mo for 90K requests; DeepSeek sells no such plan |
| Deployment regions · where it runs | AQwen | Qwen prices mainland, Singapore, US and global rows; DeepSeek's API is China-hosted |
| Price stability · rate-card churn | BDeepSeek | Qwen's headline rate is a limited-time 50% promo; DeepSeek's discount rate is the standard card |
| Best overall | ·Depends | See the decision tree below |
CHOOSE A · QWEN
If you need the smarter model or a flat coding bill.
- Reasoning — GPQA Diamond 92.4 vs 90.1 and MMLU-Pro 89.6 vs 87.5, the widest honest gaps on this page
- Coding Plan — ¥200/mo buys 90,000 requests/month in Claude Code, Cursor, Codex or Qwen Code with no per-token risk
- Lineup — Max, Plus, Flash, Coder and VL tiers let you route cheap work down without leaving the vendor
- Regions — priced deployment rows for mainland China, Singapore, the US and a global endpoint
- Version pinning — dated snapshots like qwen3.7-max-2026-05-20 keep an agent's behaviour reproducible
CHOOSE B · DEEPSEEK
If you need open weights or a much smaller bill.
- Cost — ~6× cheaper input and ~10× cheaper output for a 0.2-point SWE-bench difference
- MIT weights — V4 Pro is downloadable and self-hostable with no commercial restrictions
- Cache economics — a published $0.003625 cache-hit rate makes repeated context nearly free
- Coding parity — SWE-bench Verified 80.6 vs 80.4 means you give up nothing on agentic work
- Simple card — flat USD pricing, no region tiers and no promo clock ticking on the headline rate
| Aspect | Qwen | DeepSeek |
|---|---|---|
| API · inputper 1M tokens · from snapshot | $2.77 | $0.43 B wins |
| API · outputper 1M tokens · from snapshot | $8.31 | $0.87 B wins |
| Effective API costBlended workload $/1M · from snapshot | $3.21 | $0.14 B wins |
| Real cost / 1M charsTokenizer-adjusted prose — the tokenizer tax | $0.53 | $0.08 B wins |
| API context windowMax input tokens · from snapshot | 1M | 1M |
| Free consumer appWhat a non-paying user gets | $0 · Qwen Studio chat.qwen.ai runs free and defaults to Qwen3.7-Plus — the Max model on this page is not the app default | $0 · full app chat.deepseek.com plus iOS/Android; DeepSeek advertises free chat and publishes no consumer message caps |
| Fixed-fee coding planFlat monthly billing for agents | ¥200/mo · Coding Plan Pro 90,000 requests/month (6,000 per 5h) in Claude Code, Cursor, Codex, Qwen Code; tool use only — API calls are barred, and inputs feed model improvement. Lite is closed to new buyers A wins | None — pay-per-token No coding subscription; you meter the API, or self-host the MIT weights for flat infra cost |
| Free API trial quotaTokens before you pay | 1M tokens · mainland only Model Studio grants 1M free tokens for 90 days on the China-mainland deployment; the international, Singapore and US rows carry no free quota | Not published The DeepSeek pricing page lists no free trial allocation — assume you pay from the first token |
| Team / enterpriseSeats, admin, residency | Alibaba Cloud account Billed through Model Studio with RAM sub-accounts and a Token Plan (Team Edition); regional endpoints incl. Singapore and the US | API or self-host No per-seat tier; regulated teams run the open weights on their own infrastructure to keep data in-region |
| Capability | Qwen | DeepSeek |
|---|---|---|
| API context window | 1M tokens | 1M tokens |
| Open weights / self-host | ✗ (closed, API only) | ✓ MIT license |
| Open-weight sibling models | ✓ Qwen3.5 / 3.6 (Apache-2.0) | ✓ V4 Flash (MIT) |
| Max output tokens | ~ Not stated on price card | 384K |
| Thinking + non-thinking modes | ✓ Both modes | ✓ Both (thinking default) |
| Prompt / context caching | ~ Discount, rate not itemized | ✓ $0.003625 cache hit |
| Batch API discount | ✓ 50% off | ✗ (none published) |
| Tool calls / function calling | ✓ | ✓ |
| JSON output mode | ~ Not on price card | ✓ Documented |
| FIM completion | ~ Via Qwen Coder tier | ✓ Non-thinking mode |
| Vision / image input | ✗ (text + code; use Qwen VL) | ✗ (text-only) |
| Dated version pinning | ✓ 2026-05-20 / 06-08 snapshots | ✗ (rolling ids only) |
| Model lineup breadth | ✓ Max / Plus / Flash / Coder / VL | V4 Pro + V4 Flash |
| Regional deployment rows | ✓ CN / Singapore / US / global | ~ China-hosted API |
| OpenAI-compatible API | ✓ | ✓ |
| Anthropic-compatible endpoint | ✓ Coding Plan base URL | ~ Not published |
| Fixed-fee coding subscription | ✓ ¥200/mo Coding Plan Pro | ✗ |
| Free consumer chat app | ✓ Qwen Studio | ✓ Web + iOS/Android |
| Pricing currency | ~ CNY, region-tiered | ✓ Flat USD |
| Headline rate is promotional | ~ Limited-time 50% off | ✓ Standard card |
| Legacy alias deprecation | ~ Rolling alias re-points | ~ chat/reasoner end 2026-07-24 |
§ 04 / BENCHMARKS
The numbers, not the spin.
reasoning · GPQA Diamond
knowledge · MMLU-Pro
agentic-coding · SWE-bench Verified
A · ALIBABA
Qwen
The broader, smarter, pricier family — Alibaba's closed flagship sits on top of a full ladder of cheaper Qwen tiers and a flat-fee coding plan.
Strengths
- Reasoning — the best scores on this page: GPQA Diamond 92.4 and MMLU-Pro 89.6
- Coding Plan — ¥200/mo for 90K requests across Claude Code, Cursor, Codex and Qwen Code
- Ladder — Max, Plus, Flash, Coder and VL tiers let you route work by cost instead of switching vendor
- Regions — priced mainland, Singapore, US and global deployment rows
- Reproducibility — dated snapshots pin an agent to one model build
Weaknesses
- Closed-weight — no self-hosting, no fine-tuning of the Max model, no exit from the API
- ~6× the input cost and ~10× the output cost of DeepSeek V4 Pro
- The headline rate is a limited-time 50% promo, so budgets built on it can double
- Context-cache discount exists but the cache-read rate isn't itemized, so caching savings can't be modelled
- Coding Plan bars API use and feeds your inputs into model improvement
Best for
- Hard reasoning and research work where 2 points of GPQA matter
- Developers who want a predictable flat coding bill inside an existing IDE agent
- Teams that need a Singapore, US or global endpoint
- Workloads that route across cheap and expensive tiers of one vendor
B · DEEPSEEK
DeepSeek
The open, cheap one — MIT weights, a flat USD card, and agentic coding that matches a model costing several times more.
Strengths
- MIT weights — V4 Pro is downloadable and self-hostable with no commercial restrictions
- Cost — ~6× cheaper input and ~10× cheaper output, with no tokenizer penalty to claw any of it back
- Cache — a published $0.003625 cache-hit rate makes repeated context effectively free
- Coding parity — SWE-bench Verified 80.6 vs Qwen's 80.4
- Simplicity — one flat USD card, no region tiers, no promo clock
Weaknesses
- Trails on reasoning and knowledge — GPQA 90.1 and MMLU-Pro 87.5
- Hosted API runs in China, so regulated teams have to self-host to control residency
- No batch discount and no free trial quota published
- Thin lineup — V4 Pro and V4 Flash only, with no vision or coder-specific tier
- The deepseek-chat and deepseek-reasoner aliases retire on 2026-07-24
Best for
- High-volume API workloads where the bill dominates
- Teams that must self-host for data sovereignty or air-gapped deployment
- Agentic coding at a fraction of frontier cost
- Fine-tuning and derivative work on open weights
01
Startup shipping an agent on a tight burn rate
You're metering an agentic product on the API. Output tokens dominate the bill and a 10× rate difference decides whether the unit economics work.
Reasoning: DeepSeek is ~10× cheaper on output and ties Qwen on SWE-bench Verified (80.6 vs 80.4), so the quality you buy with the premium is roughly zero for coding-shaped work. Qwen's reasoning edge is real but 2 points of GPQA rarely pays a 6–10× multiplier at volume.
Picked
DeepSeek
Runner-up: Qwen3.7 Flash or Plus if you want Alibaba's ladder at a lower tier
02
Regulated team that cannot send data to a China-hosted API
A bank or health provider needs the model inside its own VPC, with weights it controls and an audit trail that survives a vendor change.
Reasoning: DeepSeek V4 Pro ships MIT weights on Hugging Face, so you self-host and the residency question disappears. Qwen3.7 Max is closed-weight and API-only — Alibaba's Singapore or US endpoints help with latency and region, but you're still sending data to a vendor you can't audit.
Picked
DeepSeek
Runner-up: Qwen's open-weight Qwen3.5 / 3.6 siblings if you want an Apache-2.0 self-host instead
03
Research team pushing the hardest reasoning problems
Graduate-level science questions and long analytical chains, where a wrong answer costs far more than a token.
Reasoning: Qwen3.7 Max leads GPQA Diamond 92.4 vs 90.1 and MMLU-Pro 89.6 vs 87.5. Both gaps are modest, but they're consistent and they point the same way — and on low-volume, high-stakes work the price multiple barely registers.
Picked
Qwen
Runner-up: DeepSeek V4 Pro when the same questions run at volume
04
Developer who wants a predictable monthly coding bill
You live in Claude Code or Cursor all day and hate watching a per-token meter, and you want one number on the invoice.
Reasoning: Alibaba's Coding Plan Pro is ¥200/mo for 90,000 requests, wired into Claude Code, Cursor, Codex and Qwen Code — DeepSeek sells no equivalent, so you'd meter tokens instead. Read the terms first: the plan bars scripted API use and your inputs feed model improvement.
Picked
Qwen
Runner-up: DeepSeek's API if per-token metering doesn't bother you
05
RAG pipeline replaying the same long context
You re-send a large document corpus on nearly every request, so cache behaviour drives the bill more than the headline rate.
Reasoning: DeepSeek publishes a $0.003625 cache-hit input rate against $0.435 cache-miss — roughly a 99% discount you can actually model. Alibaba says Qwen3.7 Max gets a context-cache discount but doesn't itemize the rate, which makes forecasting guesswork.
Picked
DeepSeek
Runner-up: Qwen with batch calls at 50% off if your workload is offline
06
Team standardising on one Chinese vendor across tasks
You need cheap classification, mid-tier chat, vision, and a frontier tier — and you'd rather not integrate four vendors.
Reasoning: Qwen prices Max, Plus, Flash, Coder and VL tiers on one card with one key, so you route by cost without leaving Model Studio. DeepSeek gives you V4 Pro and V4 Flash and no vision tier, so a mixed workload will eventually pull in a second vendor.
Picked
Qwen
Runner-up: DeepSeek plus an open-weight VL model for the vision leg
Frequently asked.
Common questions about this comparison, with sources where they matter.
Q · 01 Is Qwen or DeepSeek better overall? +
Neither sweeps. Qwen3.7 Max leads reasoning and knowledge — GPQA Diamond
92.4 vs 90.1, MMLU-Pro 89.6 vs 87.5 — while DeepSeek V4 Pro is MIT open-weight, roughly 6× cheaper on input and 10× cheaper on output, and ties on agentic coding (SWE-bench Verified 80.6 vs 80.4). If your work is coding or high-volume, DeepSeek's economics are hard to argue with; if it's hard reasoning at low volume, Qwen earns the premium.Q · 02 Which is cheaper? +
DeepSeek, by a wide margin — roughly 6× on input and 10× on output. The tokenizers are close enough on English prose (5.23 vs 5.26 characters per token) that the gap neither widens nor narrows much once you measure real cost per 1M characters (see the tokenizer-tax row above). Two caveats: Qwen's headline rate is currently a limited-time 50% promotion, and Qwen offers 50%-off batch calls that DeepSeek doesn't publish. Model your own mix in the LLM API cost calculator.
Q · 03 Are both Qwen and DeepSeek open source? +
Not on this matchup. DeepSeek V4 Pro ships open weights under the MIT license and is downloadable from Hugging Face. Qwen3.7 Max is closed-weight and API-only — a deliberate break from Alibaba's open-model history. Alibaba still publishes open-weight siblings (Qwen3.5 and Qwen3.6 under Apache-2.0), but those are different, weaker models than the Max flagship compared here.
Q · 04 Which is better for coding? +
It's a tie on the neutral number that matters: SWE-bench Verified
80.6 (DeepSeek) vs 80.4 (Qwen). A 0.2-point gap is noise. Since the models are level, the decision falls to cost — where DeepSeek wins outright — or to billing shape, where Alibaba's ¥200/mo Coding Plan Pro gives you a flat monthly bill inside Claude Code, Cursor or Codex.Q · 05 Do the benchmark gaps actually matter? +
Read them as small. Qwen's leads are 2.3 points on GPQA Diamond and 2.1 on MMLU-Pro — consistent in direction, but well short of a generational gap, and the two models were measured at different configurations (Qwen as-reported, DeepSeek at max effort). Treat them as a nudge toward Qwen on reasoning-heavy work, not as proof it's a different class of model.
Q · 06 Can I self-host either one? +
DeepSeek V4 Pro, yes — MIT weights, no commercial restrictions, fine-tuning and redistribution permitted. Qwen3.7 Max, no. If self-hosting is the requirement and you still want Alibaba's family, the open Qwen3.5 / Qwen3.6 releases are Apache-2.0, but you should benchmark them yourself rather than assume Max-level quality.
Q · 07 Where does my data go? +
DeepSeek's hosted API runs in China; self-hosting the open weights is the only way to fully control residency. Qwen runs on Alibaba Cloud Model Studio with separately priced mainland, Singapore, US and global deployment rows, so you can pick a region — but note that Alibaba's Coding Plan terms state your inputs and generated content are used for service and model improvement, which is a different bargain than the metered API.
Q · 08 Is there a free tier? +
Both consumer chat apps are free — Qwen Studio at chat.qwen.ai (which defaults to Qwen3.7-Plus, not the Max model priced here) and DeepSeek on web plus iOS/Android. On the API they diverge: Alibaba grants 1M free tokens for 90 days only on the China-mainland deployment — the international, Singapore and US rows get nothing — and DeepSeek's pricing page publishes no free allocation at all.
Q · 09 Can I use both? +
That's the pragmatic answer for many teams: DeepSeek for the high-volume bulk of the work, Qwen3.7 Max for the reasoning-hard minority. Both expose OpenAI-compatible endpoints, so a router (LiteLLM, OpenRouter) can switch per request. Self-hosting DeepSeek and calling Qwen for escalations is a common shape.
Q · 10 Why does this page quote CNY for Qwen and USD for DeepSeek? +
Because that's how each vendor prices. Alibaba publishes Qwen's rates in CNY per 1M tokens — the international row is
¥18.736 in / ¥56.207 out — which we convert at a dated FX rate, so the USD figure moves slightly with the exchange rate. DeepSeek publishes flat USD. See the Qwen3.7 Max pricing page for the conversion basis.