What TSK-1 Found
TSK-1 hasn't run these two head-to-head yet; here's what public benchmarks show. Qwen's open-weight models have not appeared in any TSK-1 build (Qwen is listed as available in the dataset), while DeepSeek V4 has been tested several times. DeepSeek V4 Flash built the best-looking app of its test (Aug 1, 2026), and on Aug 3, 2026 it was the cheapest, the fastest, and the cleanest all at once. Until TSK-1 runs Qwen, the comparison rests on published vendor benchmarks and licence terms, not on a real build we ran and opened ourselves.
- DeepSeek: Aug 1, 2026 — best-looking app of its test. Aug 3, 2026 — cheapest, fastest, and cleanest all at once. Aug 6, 2026 — best at holding your information (60 separate fields captured).
- Qwen: Tested in TSK-1 on Aug 25, 2026, scoring 50 out of 100 (dated findings).
See the full evidence at /tsk/deepseek, /tsk/qwen, and the TSK-1 hub.
The Distinction That Decides Everything
Qwen's open weights and Qwen's API tiers are different generations of model. Every tier whose name ends in -max, -plus, or -flash is proprietary and API-only, and the open ladder sits a generation behind it. DeepSeek does not work that way: the models it runs are the models it publishes, all under MIT.
Get that wrong and everything downstream is wrong — the licence you think you have, the model you think you can self-host, and the price you think you are paying.
TL;DR: Qwen publishes Apache 2.0 weights at Qwen3.6 (a 27B dense model and a multimodal 35B-A3B mixture-of-experts) while selling a separate API-only line that differs between international and China mainland endpoints. DeepSeek publishes one generation, V4, under MIT, at $0.44 / $1.32 per million tokens on Flash at peak. Route both inside Taskade Genesis.
What Ships Open, and What Only Ships as an API
Open, with downloadable weights under Apache 2.0:
- Qwen3.6-27B — dense.
- Qwen3.6-35B-A3B — Mixture-of-Experts, 35B total and 3B used per token, and multimodal via a vision encoder that handles both images and video.
- The Qwen3.5 ladder — 397B-A17B, 122B-A10B, 35B-A3B, 27B, 9B, 4B, 2B, and 0.8B, which is an unusually complete range of sizes.
Proprietary, API-only: qwen3.7-max, qwen3.7-plus, qwen3.6-flash, qwen3.5-omni-plus, and every other -max / -plus / -flash tier. No Qwen3.7 weights exist in the official Hugging Face organisation, so do not describe those tiers as open.
Alibaba's own wording on the open half is unambiguous: "All our open-weight models are licensed under Apache 2.0."
Region Changes the Model List
There is no single global Qwen lineup, and quoting one is the fastest way to send a reader to a model they cannot call.
| Endpoint | Top tier | Mid tier | Fast tier |
|---|---|---|---|
| International (Singapore) Model Studio | qwen3.7-max |
qwen3.7-plus |
qwen3.6-flash |
| China mainland | qwen3.8-max |
qwen3.7-plus |
qwen3.7-flash |
Mainland runs ahead of international, and mainland token prices are lower. Always name the region alongside the model.
What Runs on the DeepSeek Side in August 2026
DeepSeek moved a whole generation this summer, and a lot of published guidance is now describing endpoints that no longer answer.
- V3 and R1 are gone, and there is no R2. The current generation is V4.
deepseek-chatanddeepseek-reasonerwere retired and became inaccessible on 24 July 2026.- The two production ids are
deepseek-v4-flashanddeepseek-v4-pro, plus an experimentaldeepseek-v4-flash-vision-expon Flash's rate card. - Reasoning is a mode, not a model.
deepseek-v4-flashexposes reasoning effort atlow,high, andmax. - Parameters:
deepseek-v4-prois 1.6T total with 49B used per token.deepseek-v4-flashis roughly 284–304B total with 13B used per token — vendor sources disagree on the total, so treat it as a range.
Licences: Apache 2.0 vs MIT
| Qwen open weights | DeepSeek V4 | |
|---|---|---|
| Licence | Apache 2.0, no size exceptions | MIT |
| Covers the weights | ✅ Yes | ✅ Yes |
| Covers the repository code | ✅ Yes | ✅ Yes, explicitly |
| Commercial use | ✅ Permitted | ✅ Permitted |
| Redistribute fine-tunes | ✅ Permitted with attribution | ✅ Permitted, retain notice and state changes |
| Monthly-active-user cap | None | None |
| Applies to every model in the family | 🔴 No — only the open ladder | ✅ Yes |
Both are genuinely permissive, and both are a different universe from a closed frontier model, where self-hosting and on-prem data residency are simply not purchasable at any price. The practical difference is coverage: MIT covers everything DeepSeek ships, while Apache 2.0 covers only the half of Qwen that ships as weights.
Benchmarks: What We Can Actually Cite
The published benchmark for Qwen3.6-35B-A3B is SWE-bench Verified 73.4% and MMLU-Pro 85.2% — notable because that is a 35B-total model activating 3B parameters per token, not a frontier-scale one.
We are deliberately not printing a benchmark table pitting the two families against each other. Scores move between point releases, vendors publish under different harnesses, and a comparison page that quotes a stale leaderboard does more harm than an honest gap. Run your own evaluation on your own prompts before you commit a pipeline to either.
Price per Million Tokens
International Model Studio for Qwen (mainland is lower), published DeepSeek API rates for V4.
| Model | Input | Output |
|---|---|---|
deepseek-v4-flash — peak |
$0.44 (cache hit $0.014) | $1.32 |
deepseek-v4-flash — off-peak |
$0.22 (cache hit $0.007) | $0.66 |
deepseek-v4-pro — peak |
$1.32 (cache hit $0.044) | $3.96 |
deepseek-v4-pro — off-peak |
$0.66 (cache hit $0.022) | $1.98 |
qwen3.6-flash (intl) |
$0.25 up to 256K, $1.00 from 256K–1M | $1.50 / $4.00 |
qwen3.7-plus (intl) |
$0.40 up to 256K, $1.20 from 256K–1M | $1.60 / $4.80 |
qwen3.7-max (intl) |
$2.50 | $7.50 |
Two things to hold on to. First, Qwen's commercial tiers price in context bands — cross 256K and the input rate roughly triples on qwen3.6-flash, which changes the economics of long-document work more than the headline number suggests. Second, DeepSeek prices on a clock — off-peak is exactly half of peak, and peak covers only 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. Batch work you can schedule outside those windows costs half without changing a line of code. Every figure in this table is the rate published as of August 2026.
A third point worth its own line: on DeepSeek a cache hit is roughly 31× below a cache miss on flash and exactly 30× below on pro, at either clock rate, which is a far bigger lever than choosing between the two models. Qwen's cached-input prices for the commercial tiers are not published in a form we can cite, so no equivalent figure appears for them above.
The Routing Matrix
Pick per task, and keep the cost consequence in the same view as the capability.
| Task | Reach for | Why | Cost consequence |
|---|---|---|---|
| Cheap high-volume classification | deepseek-v4-flash |
Lowest published rate of the two families | $0.44 in / $1.32 out per 1M at peak, $0.22 / $0.66 off-peak |
| Repeated prompts over a fixed corpus | deepseek-v4-flash with cache hits |
Stable prefix bills at $0.014 per 1M at peak, $0.007 off-peak | Roughly 31x off the input side |
| Long-context reasoning, 1M tokens | deepseek-v4-pro |
1M native context, no scaling technique required | $1.32 / $3.96 per 1M at peak — a 500K prompt is a real line item |
| Vision — image or video in the loop | Qwen3.6-35B-A3B |
The multimodal model that also ships as open weights | Self-host cost, or your provider's rate |
| Agentic coding | Qwen3-Coder variants, or V4 at high effort |
Qwen keeps a dedicated coder line; DeepSeek uses one endpoint with a dial | Coder-Next ~80B is far cheaper to run than 480B-A35B |
| Multilingual content | Qwen commercial tiers, region-matched | Alibaba's commercial line is the broader multilingual bet | qwen3.7-plus $0.40 / $1.60 up to 256K, intl |
| On-prem, air-gap, data residency | Qwen3.6-35B-A3B or deepseek-v4-flash |
Apache 2.0 and MIT both permit self-hosting outright | Your GPU cost, no per-token bill |
| Smallest viable footprint | Qwen3.5 ladder down to 0.8B | Nothing else here goes that small | Runs where a frontier model cannot |
| Overnight or weekend batch runs | Either DeepSeek model, scheduled off-peak | DeepSeek halves the rate outside peak hours | 50% off input and output |
| A workflow that needs several of these | Route across models | No single family wins every row | Allowance included in the subscription |
Self-Host Reality Check
Both licences permit self-hosting. Hardware is what decides whether you actually will.
Qwen3.6-35B-A3Bis the realistic target. 35B total with only 3B used per token is the reason Qwen's open ladder shows up so often in on-prem deployments, and it is multimodal on top of that.- The Qwen3.5 ladder gives you an exit at every size — 397B-A17B at the top, 0.8B at the bottom. If a model does not fit your hardware, there is a smaller one with the same licence.
deepseek-v4-proat 1.6T total is a serious commitment.deepseek-v4-flashat roughly 284–304B total and 13B used per token is the more approachable DeepSeek self-host, but it is still an order of magnitude above the practical Qwen target.
For most teams below serious scale, a managed gateway is cheaper and simpler than running either. The licence still matters even if you never self-host, because it is what makes the option available at all when a compliance requirement lands.
Choose DeepSeek If
A comparison that only flatters one side is not worth reading. DeepSeek is the better pick when:
- You want one licence covering everything you touch. MIT applies to the code and the weights, and the models DeepSeek runs are the models it publishes. There is no open-versus-commercial line to keep track of.
- You need a million tokens of context natively. Both V4 models carry it without a scaling technique, with a 384K maximum output.
- Per-token cost is the binding constraint. On published international rates,
deepseek-v4-flashis the cheapest option on this page by a wide margin, and its cache-hit rate, at $0.014 per 1M at peak, is roughly 31x lower still.
Choose Qwen when you need vision in an open model, a specific parameter size that fits your hardware, a dedicated coding line, or a commercial multilingual tier — and when you are comfortable tracking which half of the family a given model belongs to.
The Taskade Angle: Route Across Labs, Not Between Them
Taskade routes across 15+ frontier models from OpenAI, Anthropic, and open-weight providers inside a single workspace, with the AI allowance included in the subscription instead of an API account and a key per lab. Plans are Free, Pro $10/mo billed annually, Business $25/mo billed annually, Max $100/mo billed annually, and Enterprise $250/mo billed annually.
What that changes day to day:
- Per-step model choice. Each agent or automation step picks its own model, so bulk extraction and final reasoning do not have to share a bill.
- No account sprawl. One workspace instead of a separate console, key, and invoice for every lab.
- Routing survives a repricing or a retirement. DeepSeek retired two model ids outright this summer and has announced a price rise; that is a routing change, not a migration.
- Work that outlives the prompt. Projects, agents, and automations sit alongside 100+ integrations that flow both ways — triggers pull events in from tools like Slack, Gmail, and Google Calendar, and actions push data out to Stripe, Notion, and GitHub.
Model availability is flag-driven and changes as labs ship, so treat the in-product picker rather than any comparison page as the source of truth for what you can call today.
See 10 Best Open-Source AI LLMs in 2026 for the wider open-weight field.
Final Word: Two Different Shapes of Open
Qwen gives you the deeper bench — an Apache 2.0 ladder from 0.8B to 397B, a multimodal open model, and a dedicated coding line — at the cost of tracking which models are actually open and which region you are calling. DeepSeek gives you one MIT-licensed generation where the published model and the served model are the same thing, with a native million-token context and the lowest published per-token price here.
Neither answer holds for every step of a real workflow. Route per task, and keep the price of each choice in front of you when you make it.
▲ Memory feeds Intelligence. ■ Intelligence triggers Execution. ● Execution creates Memory. Two open-weight families. One workspace.
This is the origin of living software. 🌱
Build with Qwen and DeepSeek in one workspace →
Related reading
- 10 Best Open-Source AI LLMs in 2026 — The full open-weight ranking.
- DeepSeek vs ChatGPT — Token pricing against a per-seat consumer product.
- Kimi vs DeepSeek — The other open-weight head-to-head.
- Multi-Model AI Access — How Taskade routes 15+ models.
- Model Credits — Per-model credit costs and plan quotas.
- Taskade MCP Server — Use Claude Desktop or Cursor with your workspace.
- Free Claude Alternative — How premium frontier compares.
- Free ChatGPT Alternative — The OpenAI side.
- TSK-1 DeepSeek profile — Full benchmark evidence for the DeepSeek family.
- TSK-1 Qwen profile — Full benchmark evidence for the Qwen family.
- TSK-1 hub — The complete model benchmark dataset.
