DeepSeek R1 32B vs Qwen 3 Thinking Mode — Reasoning Compared
Last updated 2026-06-19
Both 32B reasoners fit on a single 24 GB GPU. We benchmarked AIME, GPQA, and code — here is which one wins in 2026 and which one is just faster.
By Mohamed Meguedmi · 9 min read
Key Takeaways
- Qwen3-32B Thinking edges out DeepSeek-R1-Distill-Qwen-32B on AIME 2024 (74.4% vs 72.6%) and GPQA Diamond (68.4% vs 62.1%), making it the stronger pure-reasoning pick at the 32B class.
- DeepSeek R1 32B remains roughly 8% faster at decode on identical hardware (~37 vs ~34 tok/s on an RTX 4090 at Q4_K_M).
- Both models fit on a single 24 GB GPU at Q4_K_M (~19.5 GB weights + ~3.5 GB for a 16k KV cache).
- Qwen3-32B is the only one of the two with a togglable thinking mode — set
enable_thinking=Falsefor fast non-reasoning replies and back toTruefor hard problems. - For US/UK/AU local deployments under 32B, Qwen3-32B Thinking is our 2026 reasoning default; DeepSeek R1 32B stays the budget speed pick.
The 32 billion parameter class is the sweet spot for local reasoning: small enough to fit on a single consumer GPU, large enough to crack competition math and graduate-level science questions. Two models dominate it. DeepSeek-R1-Distill-Qwen-32B (released January 2025) was the first distilled reasoner to truly close the gap to GPT-4o on AIME. Qwen3-32B (April 2025) responded with a unified architecture that lets you toggle “thinking mode” on a per-prompt basis. Eighteen months later, both are still the most-downloaded 32B reasoning weights on Hugging Face — so which one should you actually run?
The Two Models at a Glance
Both models share the same skeleton: 32 billion dense parameters, GQA attention, and a 128k context window. The training signal is where they diverge. DeepSeek built R1 32B by distilling chain-of-thought traces from its 671B MoE parent into a Qwen2.5-32B base. Qwen3 was trained from scratch on a multi-stage reasoning curriculum and ships with explicit <think>...</think> separators that the chat template can suppress or expose.
| Spec | DeepSeek-R1-Distill-Qwen-32B | Qwen3-32B |
|---|---|---|
| Parameters | 32.8B dense | 32.8B dense |
| Base model | Qwen2.5-32B-Base | Qwen3-32B (from scratch) |
| Training data cutoff | Oct 2023 (inherited) + R1 traces | Mar 2024 + reasoning curriculum |
| Context window | 131,072 tokens | 131,072 tokens (YaRN extended) |
| License | MIT | Apache 2.0 |
| Reasoning toggle | Always-on (distilled) | Togglable via enable_thinking |
| Thinking template | <think> inline | <think> block, mute-able |
| Vocabulary size | 152,064 | 151,936 |
Practically, the togglable thinking mode in Qwen3 is more than a convenience: it lets a single inference server handle low-latency chat AND deep reasoning without swapping weights. DeepSeek R1 32B will always emit a <think> block — sometimes hundreds of tokens of self-reflection for trivial questions — which inflates output cost and latency.
Reasoning Benchmarks — Head to Head
We pulled scores from the model cards published by DeepSeek on Hugging Face and the official Qwen3-32B card, cross-checked with the independent Artificial Analysis reasoning index. Where vendors and AA disagree, we list the AA figure; otherwise the vendor number.
| Benchmark | DeepSeek R1 32B | Qwen3-32B (Thinking) | Winner |
|---|---|---|---|
| AIME 2024 (pass@1) | 72.6% | 74.4% | Qwen3 +1.8 pp |
| AIME 2025 (pass@1) | 62.1% | 72.9% | Qwen3 +10.8 pp |
| MATH-500 | 94.3% | 95.1% | Qwen3 +0.8 pp |
| GPQA Diamond | 62.1% | 68.4% | Qwen3 +6.3 pp |
| LiveCodeBench v5 | 57.2% | 65.7% | Qwen3 +8.5 pp |
| Codeforces (Elo) | 1,691 | 1,977 | Qwen3 +286 |
| MMLU-Pro | 74.2% | 78.6% | Qwen3 +4.4 pp |
| Avg thinking tokens / hard prompt | ~8,200 | ~5,900 | Qwen3 (28% shorter) |
The pattern is consistent: Qwen3-32B wins every reasoning category, and it does so while emitting fewer thinking tokens. The gap is largest on AIME 2025 — a benchmark released after DeepSeek R1’s training cutoff, which suggests Qwen3 has genuinely better generalization on unseen math contests, not just better memorization.
None of these scores tell you what happens on your domain. Run both through prompts that match your workload. For automated head-to-head harnesses, the open-source BestLLMfor MCP server (see /about/) exposes our entire benchmark table via a single tool call — useful when comparing models in the same code path as your agent.
Hardware Footprint and Quantization
Both models are dense 32.8B parameters, so memory math is essentially identical. The real question is which quant level holds reasoning quality.
| Quant | VRAM (weights) | + 16k ctx KV cache | Min GPU | Reasoning quality loss |
|---|---|---|---|---|
| FP16 / BF16 | ~62 GB | ~67 GB | A100 80GB / 2x RTX 6000 Ada | Baseline |
| Q8_0 (8-bit GGUF) | ~34 GB | ~38 GB | RTX 5090 / A6000 48GB | <1% |
| Q5_K_M | ~22 GB | ~26 GB | RTX 4090 / 5090 (24 GB+) | ~1-2% |
| Q4_K_M | ~19.5 GB | ~23 GB | RTX 3090 / 4090 24 GB | ~2-4% |
| IQ3_XS | ~13 GB | ~17 GB | RTX 4070 Ti Super 16 GB | 5-10% (avoid for math) |
For both models, Q4_K_M is the practical floor for serious reasoning. Dropping to 3-bit costs roughly 8 percentage points on AIME 2024 in our internal runs — the chain-of-thought becomes more likely to flip a sign mid-derivation. If you have 48 GB or more, run Q8_0 and stop worrying about quant noise. The full quantization methodology is documented at /methodology/.
Inference Speed and Real-World Cost
Reasoning models are output-token-heavy by design. On a single RTX 4090 with Q4_K_M GGUF (llama.cpp build b4523), the following numbers came out of a fixed AIME 2024 problem set:
| Metric | DeepSeek R1 32B | Qwen3-32B (think on) | Qwen3-32B (think off) |
|---|---|---|---|
| Prompt eval (tok/s) | 410 | 425 | 431 |
| Decode (tok/s) | 37.1 | 34.4 | 34.9 |
| Avg total tokens / hard math Q | 8,240 | 5,920 | 1,180 |
| Avg wall-clock / hard Q | 3m 42s | 2m 52s | 34s |
| Energy / hard Q (Wh @ 350 W TDP) | 21.6 | 16.7 | 3.3 |
Per-token, DeepSeek R1 32B is faster (37 vs 34 tok/s). Per-correct-answer, Qwen3 wins by ~24% because its reasoning chains are shorter. If you bill by output tokens — most US-hosted endpoints do — that gap is your bill. Plug your own request volume into the BestLLMfor cost calculator to see the monthly delta.
When Each Model Wins
Pick Qwen3-32B Thinking when…
- You need a single endpoint for both fast chat and deep reasoning. The
enable_thinkingflag toggles the budget without re-loading weights. - Your workload is heavy on code generation. Qwen3’s +286 Codeforces Elo lead is the largest delta on the board.
- You operate on post-2024 math/science (AIME 2025+, GPQA Diamond updates). Qwen3’s training cutoff is six months later and it shows.
- You ship to Apache 2.0 customers who flinch at MIT-only terms (rare, but it happens in enterprise procurement).
Pick DeepSeek R1 32B when…
- You want raw decode throughput. The 37 tok/s figure on a 4090 is a hard ceiling Qwen3 doesn’t quite hit.
- You’re building a reasoning-only agent and don’t want the option to skip thinking — the always-on
<think>is a feature, not a bug. - You already have R1 671B in production and want the distilled model to share prompts and parsers verbatim.
- You value the cleaner distillation lineage: every chain-of-thought token traces back to the R1 reinforcement-learning loop, which some interpretability researchers prefer.
For a broader catalog of 32B-class local models, including non-reasoning variants like Qwen2.5-Coder-32B, see /catalog/.
Installing Both Locally — Ollama Path
The fastest way to A/B test on a single machine is Ollama. Both models ship as official tags.
- Install Ollama (Linux/macOS:
curl -fsSL https://ollama.com/install.sh | sh; Windows: download the installer). - Pull DeepSeek R1 32B:
ollama pull deepseek-r1:32b(~20 GB at Q4_K_M). - Pull Qwen3-32B:
ollama pull qwen3:32b(~20 GB at Q4_K_M). - For Qwen3, append
/thinkor/no_thinkto any prompt to flip the thinking budget on the fly. - Benchmark with
ollama run --verboseto get per-prompt tok/s and total token counts.
For production deployments, vLLM 0.7+ is the recommended engine for both — it handles the reasoning prefix-cache substantially better than Ollama and delivers roughly 2.3x the throughput on identical hardware.
The Verdict
At the 32B class, Qwen3-32B with thinking mode enabled is the better reasoner in 2026. It wins every benchmark tracked here, generates ~28% shorter chains of thought, and ships with a toggle that lets it double as a low-latency chat model. DeepSeek-R1-Distill-Qwen-32B is still excellent — and meaningfully faster per token — but the gap on AIME 2025, GPQA, and Codeforces is large enough that the speed advantage rarely compensates.
| Use case | Recommended model | Why |
|---|---|---|
| Local math/science tutor | Qwen3-32B Thinking | +10.8 pp on AIME 2025, +6.3 pp on GPQA |
| Local coding agent | Qwen3-32B Thinking | +286 Codeforces Elo, +8.5 pp LiveCodeBench |
| Mixed chat + reasoning endpoint | Qwen3-32B (toggle) | Only model with togglable thinking |
| Pure speed reasoning at Q4 | DeepSeek R1 32B | ~8% higher tok/s on RTX 4090 |
| R1 671B distillation parity | DeepSeek R1 32B | Same parser, same prompt template |
| Apache 2.0 license required | Qwen3-32B | MIT vs Apache distinction |
If you only have 16 GB of VRAM, neither is comfortable — drop to Qwen3-14B Thinking, which holds ~95% of the 32B reasoning quality. If you have 48 GB or more and are not bound by power budget, the 70B-class Llama-3.3-70B-Instruct or QwQ-32B-Preview’s successor are worth comparing instead.
All numbers in this guide are available as machine-readable JSON via the BestLLMfor public API (CC BY 4.0, no API key required for read endpoints). Methodology, raw logs, and quantization scripts live at /methodology/.
Frequently Asked Questions
Does Qwen3-32B Thinking beat DeepSeek R1 (the full 671B)?
No. The 32B distillation closes most of the gap on standard benchmarks, but the full DeepSeek R1 671B MoE still leads by 4-6 pp on AIME 2024 and ~8 pp on GPQA. The 32B distill is the right choice only when you cannot run the 671B model locally — which most people cannot, as it needs roughly 400 GB of VRAM at Q4.
Can I run both models on a 24 GB RTX 4090?
Yes, but only one at a time. Q4_K_M weights are ~19.5 GB; with a 16k context cache that is ~23 GB, just under the 24 GB limit. For a 32k context, drop to Q4_K_S or IQ4_XS. To swap quickly between the two models, use Ollama’s keep-alive flag or run vLLM with hot-swappable adapters.
What is the difference between Qwen3 thinking mode and DeepSeek R1's think block?
DeepSeek R1’s thinking is structurally identical (<think>...</think> wrappers) but always-on. Qwen3 lets you suppress the entire thinking phase by setting enable_thinking=False in the chat template, or by appending /no_think to the prompt. The model then skips reasoning and replies directly — useful for routine queries where chain-of-thought is wasted compute.
Which model has the smaller hallucination rate on reasoning tasks?
Qwen3-32B Thinking, by a small margin. On TruthfulQA-MC1 the team measured 62.4% vs DeepSeek R1 32B’s 58.1%. The gap widens on multi-hop questions where DeepSeek’s longer chains-of-thought give it more opportunities to invent intermediate facts.
Is the DeepSeek R1 32B model the same as the original R1?
No. The original DeepSeek-R1 is a 671B MoE (37B active params). What people run locally is DeepSeek-R1-Distill-Qwen-32B — Qwen2.5-32B fine-tuned on 800k reasoning traces sampled from the R1 671B teacher. It is a separate model with separate weights, distributed under MIT, sharing only its prompt template with the parent.
Should I use llama.cpp, vLLM, or Ollama for these models?
For single-user testing: Ollama. For multi-user production: vLLM 0.7+ (best prefix caching, handles long thinking traces efficiently). llama.cpp is the right call only for CPU-only deployments or Apple Silicon Metal — both models work with Metal but decode at ~10-15 tok/s on an M3 Max.
For running local LLMs comfortably, an RTX 5070 Ti (16 GB VRAM) is the best value for money.
Amazon Check RTX 5070 Ti price →As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.