Llama 3.2 3B Instruct vs Qwen3.5 397B A17B
Qwen3.5 397B A17B leads Llama 3.2 3B Instruct on the Artificial Analysis Intelligence Index, 21.4 to 5.7, and Llama 3.2 3B Instruct costs 7.0x less per output token ($0.3325 vs $2.34 per million). Updated October 22, 2026; 29 benchmarks have a result for both.
Llama 3.2 3B Instruct or Qwen3.5 397B A17B: which to pick
Neither model has a coding score yet, so the two cannot be ranked on coding.
Pick Llama 3.2 3B Instruct for output tokens 7.0x cheaper ($0.3325 vs $2.34 per million) and work measured by LMArena Text: coding or LMArena Text: software and IT, where it leads.
- Shared results
- 29
- Llama 3.2 3B Instruct only
- 6
- Qwen3.5 397B A17B only
- 36
- Like-for-like categories
- 7 of 9
Only benchmarks that tested both models count toward a pick; a result only one model has is shown but not compared.
Which one for your work
Each card is decided by one number both models carry. Where only one has it, the card says so and makes no call.
Coding
Vibe Coding Index, six public coding benchmarks
No call
Neither has a coding score yet.
One side not measured
Intelligence
AA Intelligence Index
Qwen3.5 397B A17B
Qwen3.5 397B A17B leads, 21.4 against 5.7.
Same board, both measured
Web design
Design Arena: websites
No call
Only Qwen3.5 397B A17B has a Design Arena: websites result, 1190.
Directional only
Long documents
Listed context window
Qwen3.5 397B A17B
Qwen3.5 397B A17B takes 262K tokens against 131K.
Listed by OpenRouter
Everyday cost
1K input + 500 output tokens
Llama 3.2 3B Instruct
Llama 3.2 3B Instruct $0.0002, Qwen3.5 397B A17B $0.0016 per turn.
List prices
At a glance
- Coding index
- Llama 3.2 3B Instruct not benchmarked, Qwen3.5 397B A17B not benchmarked.
- Widest category
- Agentic work, 74.6 points apart on average over 1 shared benchmark.
- Widest single gap
- LMArena Text: Chinese: Llama 3.2 3B Instruct 1017 against Qwen3.5 397B A17B 1498, 80.2 points apart on a 0 to 100 scale.
- Output price
- Llama 3.2 3B Instruct $0.3325 against Qwen3.5 397B A17B $2.34 per million output tokens, 7.0x apart.
- Context window
- Llama 3.2 3B Instruct takes 131K tokens, Qwen3.5 397B A17B 262K.
Benchmark by benchmark
First the score: boards each model leads, per category, where both have a result. Then every board, a category to a fold; bars are each result on a 0 to 100 scale, so compare the two in a row, not one row with another.
| Category | Llama 3.2 3B Instruct | Bars | Qwen3.5 397B A17B | Ties |
|---|---|---|---|---|
| Coding | 0 | 2 | ||
| Intelligence | 0 | 1 | ||
| Agentic work | 0 | 1 | ||
| Reasoning | 0 | 5 | ||
| Math | 0 | 2 | ||
| Knowledge | 0 | 4 | ||
| General | 0 | 14 | ||
| All shared boards | 0 | 29 |
CodingLike-for-likeSix of these boards feed the Vibe Coding Index02
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| LMArena WebDev | Not measured | 1400 | Only Qwen3.5 397B A17B measured |
| LiveCodeBench (Artificial Analysis run) | 8.3% | Not measured | Only Llama 3.2 3B Instruct measured |
| SciCode | Not measured | 44.8% | Only Qwen3.5 397B A17B measured |
| Terminal-Bench Hard (Artificial Analysis run) | Not measured | 40.9% | Only Qwen3.5 397B A17B measured |
| ALE-Bench | Not measured | 621.9 | Only Qwen3.5 397B A17B measured |
| LMArena Text: coding | 1098 | 1464 | Qwen3.5 397B A17B |
| LMArena Text: software and IT | 1103 | 1463 | Qwen3.5 397B A17B |
| LMArena WebDev: React | Not measured | 1388 | Only Qwen3.5 397B A17B measured |
| LMArena WebDev: HTML | Not measured | 1410 | Only Qwen3.5 397B A17B measured |
| Neither model has a result on the other 30 boards here | |||
IntelligenceLike-for-likePublic boards, shown as published, kept out of the index01
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| AA Intelligence Index | 5.7 | 21.4 | Qwen3.5 397B A17B |
| Epoch Capabilities Index | Not measured | 146.8 | Only Qwen3.5 397B A17B measured |
Agentic workLike-for-likePublic boards, shown as published, kept out of the index01
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| τ²-Bench Telecom | 21.1% | 95.6% | Qwen3.5 397B A17B |
| APEX-Agents | Not measured | 24.9% | Only Qwen3.5 397B A17B measured |
| BALROG | 10.1% | Not measured | Only Llama 3.2 3B Instruct measured |
| Berkeley Function Calling Leaderboard (BFCL) | 21.9% | Not measured | Only Llama 3.2 3B Instruct measured |
| Neither model has a result on the other 37 boards here | |||
ReasoningLike-for-likePublic boards, shown as published, kept out of the index05
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| GPQA Diamond (Artificial Analysis run) | 25.5% | 89.3% | Qwen3.5 397B A17B |
| Humanity's Last Exam (Artificial Analysis run) | 5.3% | 29.0% | Qwen3.5 397B A17B |
| GPQA Diamond (Epoch AI run) | Not measured | 86.4% | Only Qwen3.5 397B A17B measured |
| Mystery game puzzles | Not measured | 18.0% | Only Qwen3.5 397B A17B measured |
| LMArena Text: hard prompts | 1095 | 1448 | Qwen3.5 397B A17B |
| LMArena Text: expert | 1090 | 1462 | Qwen3.5 397B A17B |
| LMArena Text: hard prompts in English | 1141 | 1450 | Qwen3.5 397B A17B |
| LMArena Vision: diagrams | Not measured | 1277 | Only Qwen3.5 397B A17B measured |
| LMArena Vision: homework | Not measured | 1292 | Only Qwen3.5 397B A17B measured |
| Neither model has a result on the other 13 boards here | |||
MathLike-for-likePublic boards, shown as published, kept out of the index02
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| AA Math Index | 3.3 | Not measured | Only Llama 3.2 3B Instruct measured |
| AIME 2025 (Artificial Analysis run) | 3.3% | Not measured | Only Llama 3.2 3B Instruct measured |
| OTIS Mock AIME | Not measured | 88.9% | Only Qwen3.5 397B A17B measured |
| FrontierMath, tiers 1 to 3 | Not measured | 31.2% | Only Qwen3.5 397B A17B measured |
| LMArena Text: math | 1126 | 1453 | Qwen3.5 397B A17B |
| LMArena Text: mathematical work | 1128 | 1454 | Qwen3.5 397B A17B |
| Neither model has a result on the other 7 boards here | |||
KnowledgeLike-for-likePublic boards, shown as published, kept out of the index04
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| MMLU Pro (Artificial Analysis run) | 34.7% | Not measured | Only Llama 3.2 3B Instruct measured |
| LMArena Text: science | 1106 | 1457 | Qwen3.5 397B A17B |
| LMArena Text: medicine and healthcare | 1104 | 1455 | Qwen3.5 397B A17B |
| LMArena Text: legal and government | 1115 | 1441 | Qwen3.5 397B A17B |
| LMArena Text: business and finance | 1074 | 1435 | Qwen3.5 397B A17B |
| Neither model has a result on the other 13 boards here | |||
Web designDirectional onlyPublic boards, shown as published, kept out of the index7 boards, none shared
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| Design Arena: websites | Not measured | 1190 | Only Qwen3.5 397B A17B measured |
| Design Arena: UI components | Not measured | 1160 | Only Qwen3.5 397B A17B measured |
| Design Arena: game dev | Not measured | 1140 | Only Qwen3.5 397B A17B measured |
| Design Arena: data viz | Not measured | 1175 | Only Qwen3.5 397B A17B measured |
| Design Arena: 3D | Not measured | 1160 | Only Qwen3.5 397B A17B measured |
| Design Arena: all code categories | Not measured | 1179 | Only Qwen3.5 397B A17B measured |
| Design Arena: SVG | Not measured | 1139 | Only Qwen3.5 397B A17B measured |
| Neither model has a result on the other 2 boards here | |||
GeneralLike-for-likePublic boards, shown as published, kept out of the index014
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| IFBench | 26.2% | 78.8% | Qwen3.5 397B A17B |
| Long context reasoning (AA-LCR) | 4.3% | 77.3% | Qwen3.5 397B A17B |
| DTBench | Not measured | 87.5% | Only Qwen3.5 397B A17B measured |
| LMCA | Not measured | 37.9% | Only Qwen3.5 397B A17B measured |
| LMArena Text | 1110 | 1437 | Qwen3.5 397B A17B |
| LMArena Text: instruction following | 1089 | 1424 | Qwen3.5 397B A17B |
| LMArena Text: creative writing | 1094 | 1402 | Qwen3.5 397B A17B |
| LMArena Vision | Not measured | 1262 | Only Qwen3.5 397B A17B measured |
| LMArena Text: English | 1180 | 1443 | Qwen3.5 397B A17B |
| LMArena Text: non-English | 1019 | 1430 | Qwen3.5 397B A17B |
| LMArena Text: Chinese | 1017 | 1498 | Qwen3.5 397B A17B |
| LMArena Text: German | 1056 | 1445 | Qwen3.5 397B A17B |
| LMArena Text: Japanese | Not measured | 1429 | Only Qwen3.5 397B A17B measured |
| LMArena Text: Korean | Not measured | 1384 | Only Qwen3.5 397B A17B measured |
| LMArena Text: French | Not measured | 1462 | Only Qwen3.5 397B A17B measured |
| LMArena Text: Spanish | Not measured | 1441 | Only Qwen3.5 397B A17B measured |
| LMArena Text: Russian | 949 | 1430 | Qwen3.5 397B A17B |
| LMArena Text: Polish | Not measured | 1447 | Only Qwen3.5 397B A17B measured |
| LMArena Text: multi-turn | 1105 | 1446 | Qwen3.5 397B A17B |
| LMArena Text: longer queries | 1100 | 1441 | Qwen3.5 397B A17B |
| LMArena Text: writing and language | 1096 | 1415 | Qwen3.5 397B A17B |
| LMArena Text: entertainment and media | 1083 | 1399 | Qwen3.5 397B A17B |
| LMArena Vision: English | Not measured | 1268 | Only Qwen3.5 397B A17B measured |
| LMArena Vision: Chinese | Not measured | 1314 | Only Qwen3.5 397B A17B measured |
| LMArena Vision: reading text | Not measured | 1271 | Only Qwen3.5 397B A17B measured |
| LMArena Vision: humor | Not measured | 1262 | Only Qwen3.5 397B A17B measured |
| LMArena Vision: captioning | Not measured | 1230 | Only Qwen3.5 397B A17B measured |
| LMArena Vision: recognizing things | Not measured | 1265 | Only Qwen3.5 397B A17B measured |
| LMArena Vision: creative writing | Not measured | 1260 | Only Qwen3.5 397B A17B measured |
| Neither model has a result on the other 17 boards here | |||
Image generationNot comparablePublic boards, shown as published, kept out of the index0 boards, none shared
| Benchmark | Llama 3.2 3B Instruct | Qwen3.5 397B A17B | Leader |
|---|---|---|---|
| Neither model has a result on 5 boards here | |||
What each workload costs
Three fixed token mixes at each model's listed price, so per-token rates become a bill.
Chat turn
1K input + 500 output tokens
Llama 3.2 3B Instruct costs less
Repository review
50K input + 3K output tokens
Llama 3.2 3B Instruct costs less
Agent loop
200K input + 10K output tokens
Llama 3.2 3B Instruct costs less
Listed OpenRouter rates, without cache or batch discounts. A workload that does not fit a model's context window would need splitting.
Specification differences
The leading value in a row is marked; a row with a missing value names no leader. Our two rows are marked in red.
| Specification | Meta Llama 3.2 3B Instruct | Qwen3.5 397B A17B |
|---|---|---|
| Maker | Meta Llama | Qwen |
| Weights | Open weights | Open weights |
| Released | September 25, 2024 | February 16, 2026 |
| Context window | 131K | 262K |
| Input $/Mtok | $0.0505 | $0.39 |
| Output $/Mtok | $0.3325 | $2.34 |
| Coding index | Not benchmarked | Not benchmarked |
| Independent measurements behind it | 0 | 1 |
| Cost per solved task | Not computable | Not computable |
| Source | OpenRouter listing | OpenRouter listing |
Questions
Is Llama 3.2 3B Instruct better than Qwen3.5 397B A17B for coding?
Neither model has a coding score yet, so the two cannot be ranked on coding. Pick Llama 3.2 3B Instruct for output tokens 7.0x cheaper ($0.3325 vs $2.34 per million) and work measured by LMArena Text: coding or LMArena Text: software and IT, where it leads.
Which is cheaper, Llama 3.2 3B Instruct or Qwen3.5 397B A17B?
Llama 3.2 3B Instruct costs $0.0505 per million input tokens and $0.3325 per million output tokens; Qwen3.5 397B A17B costs $0.39 and $2.34. Cost per solved task is known for neither of them.
Which has the bigger context window, Llama 3.2 3B Instruct or Qwen3.5 397B A17B?
Qwen3.5 397B A17B, with 262K tokens against 131K.
How many benchmarks is the Llama 3.2 3B Instruct vs Qwen3.5 397B A17B comparison based on?
29 benchmarks have a result for both models as of October 22, 2026. Llama 3.2 3B Instruct's coding index rests on 0 independent benchmarks and Qwen3.5 397B A17B's on 1. Where only one model has a result, the page says so instead of naming a winner.