Qwen3 8B vs Llama 3.1 8B for local chat
The two most-searched 8B chat models for an 8 to 16 GB computer
Facts from the reviewed catalog
| Specification | Qwen3 8B | Llama 3.1 8B Instruct |
|---|---|---|
| Task | Text & chat | Text & chat |
| Parameters | 8.2B | 8B |
| License | Apache 2.0 | Publisher terms: Llama 3.1 |
| Commercial use | Allowedlicense text | Allowed with conditionsOver 700M monthly users needs a Meta license; show "Built with Llama" · license text |
| Quality · LMArena 2026-10-22 · text | Not on the LMArena text leaderboard of 2026-10-22 · source | 1,211 ± 41,207–1,215 · 49,605 votes · rank 345 · source |
| Memory | Q4 planning estimate (~7 GB of weights before the machine's reserve) | Q4 planning estimate (~6 GB of weights before the machine's reserve) |
| Checkpoint | official card | official card |
Only Llama 3.1 8B Instruct has a result in LMArena 2026-10-22 · text (1,211 ± 4); Qwen3 8B has none, so the two are not ranked from this source and no score is inferred for it. Blind pairwise votes by people on the hosted model, CC BY 4.0. A rating is relative; overlapping intervals mean the order is not settled. Scale: Arena rating (style control, overall) ± half the confidence interval.
Memory on three machines
Each machine's own reserve is added (graphics cards 2 GB; unified and system memory 25%, 2 to 8 GB). Planning estimates, not measured runs.
| Machine | Qwen3 8B | Llama 3.1 8B Instruct |
|---|---|---|
| RTX 4060 · 8 GB VRAM | ~9 GBIn memory, outside the fast part · offload to PC RAM (32 GB assumed) | ~8 GBFast memory holds it |
| RTX 3060 · 12 GB VRAM | ~9 GBFast memory holds it | ~8 GBFast memory holds it |
| MacBook Air 13-inch M4 · 16 GB unified | ~11 GBFast memory holds it | ~10 GBFast memory holds it |
Qwen3 8B: Q4 planning estimate (~7 GB of weights before the machine's reserve). Llama 3.1 8B Instruct: Q4 planning estimate (~6 GB of weights before the machine's reserve). Formulas
Everything that runs on each machine: RTX 4060 · 8 GB VRAM · RTX 3060 · 12 GB VRAM · MacBook Air 13-inch M4 · 16 GB unified
Which to choose
Nearly the same size, different terms. Qwen3 8B is a little larger (8.2 billion parameters against 8), which in the table below puts it just outside an 8 GB card's fast memory while Llama 3.1 8B stays inside; from 12 GB up both run well. Qwen3 8B is Apache 2.0 with no conditions; Llama 3.1 8B uses Meta's licence, which asks for a separate licence above 700 million monthly users and a “Built with Llama” notice. Llama 3.1 publishes a 128,000-token context against 32,768 for Qwen3 8B. Only Llama has a rating in the LMArena text release, so quality cannot be compared from this source.
Files and run commands
The default file of the most downloaded collected source, or the commands each card documents. Not run by this site.
Qwen3 8B
Qwen/Qwen3-8B-GGUF · Qwen3-8B-Q4_K_M.gguf · Q4_K_M · 5 GB · Qwen
ollama run hf.co/Qwen/Qwen3-8B-GGUF:Q4_K_Mllama-server -hf Qwen/Qwen3-8B-GGUF:Q4_K_Mlms get https://huggingface.co/Qwen/Qwen3-8B-GGUF- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Llama 3.1 8B Instruct
bartowski/Meta-Llama-3.1-8B-Instruct-GGUF · Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf · Q4_K_M · 4.9 GB · bartowski
ollama run hf.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_Mllama-server -hf bartowski/Meta-Llama-3.1-8B-Instruct-GGUF:Q4_K_Mlms get https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF- mlx-lm: mlx-lm runs MLX repositories; pick the MLX version on a Mac
Numbers each publisher reports
Each card's own results with its own settings. Publishers choose different tests and settings, so these values are never compared across publishers here.
Qwen3 8B · reported by Alibaba
Llama 3.1 8B Instruct · reported by Meta
- MMLU
- 69.4Llama 3.1 8B Instruct · Instruction tuned models · General
- MMLU (CoT)
- 73.0Llama 3.1 8B Instruct · Instruction tuned models · General
- MMLU-Pro (CoT)
- 48.3Llama 3.1 8B Instruct · Instruction tuned models · General
- IFEval
- 80.4Llama 3.1 8B Instruct · Instruction tuned models · General
- ARC-C
- 83.4Llama 3.1 8B Instruct · Instruction tuned models · Reasoning
- GPQA
- 30.4Llama 3.1 8B Instruct · Instruction tuned models · Reasoning
- HumanEval
- 72.6Llama 3.1 8B Instruct · Instruction tuned models · Code
- MBPP ++ base version
- 72.8Llama 3.1 8B Instruct · Instruction tuned models · Code
Every reported result
- Multipl-E HumanEval
- 50.8Llama 3.1 8B Instruct · Instruction tuned models · Code
- Multipl-E MBPP
- 52.4Llama 3.1 8B Instruct · Instruction tuned models · Code
- GSM-8K (CoT)
- 84.5Llama 3.1 8B Instruct · Instruction tuned models · Math
- MATH (CoT)
- 51.9Llama 3.1 8B Instruct · Instruction tuned models · Math
- API-Bank
- 82.6Llama 3.1 8B Instruct · Instruction tuned models · Tool Use
- BFCL
- 76.1Llama 3.1 8B Instruct · Instruction tuned models · Tool Use
- Gorilla Benchmark API Bench
- 8.2Llama 3.1 8B Instruct · Instruction tuned models · Tool Use
- Nexus (0-shot)
- 38.5Llama 3.1 8B Instruct · Instruction tuned models · Tool Use · 0-shot
- Multilingual MGSM (CoT)
- 68.9Llama 3.1 8B Instruct · Instruction tuned models · Multilingual
- MMLU · Portuguese (5-shot, macro_avg/acc)
- 62.12Llama 3.1 8B · Multilingual benchmarks · General · 5-shot
- MMLU · Spanish (5-shot, macro_avg/acc)
- 62.45Llama 3.1 8B · Multilingual benchmarks · General · 5-shot
- MMLU · Italian (5-shot, macro_avg/acc)
- 61.63Llama 3.1 8B · Multilingual benchmarks · General · 5-shot
- MMLU · German (5-shot, macro_avg/acc)
- 60.59Llama 3.1 8B · Multilingual benchmarks · General · 5-shot
- MMLU · French (5-shot, macro_avg/acc)
- 62.34Llama 3.1 8B · Multilingual benchmarks · General · 5-shot
- MMLU · Hindi (5-shot, macro_avg/acc)
- 50.88Llama 3.1 8B · Multilingual benchmarks · General · 5-shot
- MMLU · Thai (5-shot, macro_avg/acc)
- 50.32Llama 3.1 8B · Multilingual benchmarks · General · 5-shot
Values as printed on the card, with the publisher's own settings (card section: Benchmark scores › Instruction tuned models; Benchmark scores › Instruction tuned models › Multilingual benchmarks). Not comparable across publishers and not used in any ranking. Model card
Limits and dates
The published quality benchmark does not test a local quantized build. Memory estimates are not speed measurements or guaranteed hardware fits. Repository size, when available, is not VRAM. Formulas
Page updated · Catalog reviewed · LMArena 2026-10-22 · text