Mixtral 8x7B
A French startup's open-weight model that activates only 13B of its 47B parameters for any given input — a design that delivers the quality of a 70B model at the inference cost of a 13B one. Released via a magnet link with no documentation, it became the first open-weight Mixture-of-Experts model competitive with proprietary competitors.
How are Intelligence, Speed & Cost bucketed?
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
Mixtral's release proved that sparse architectures could ship in open weights at frontier-comparable quality. Every subsequent MoE open release (DeepSeek V2/V3, Qwen-MoE, Snowflake Arctic) traces its strategy to this launch. Marked the transition of open-source LLM development from "Llama clones" to architectural diversity.
Core Capabilities
Context Window
Availability
Pricing Model
Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
What it feels like
- First major MoE open-weights release that actually worked for production — 46.7B total / 13B active per token
- 6x faster inference than Llama 2 70B at equal-or-better quality on most benchmarks
- Outperformed GPT-3.5 on the LMSYS arena while being fully open-weight under Apache 2.0
- MT-Bench score earned 'best open-weights model as of December 2023'
- Vastly better than Llama 2 70B on math, code, and multilingual benchmarks
- Set the architectural template (router + sparse experts) followed by every later open MoE
Best use cases
- Open-source production inference where Llama 2 70B was too slow
- Multilingual workloads (Mixtral was particularly strong outside English at release)
- Self-hosted RAG and code-completion stacks under Apache 2.0
- Research into MoE routing, distillation, and serving
Tools to try
Not ideal for
- Frontier reasoning by 2025 — superseded by Mixtral 8x22B, Llama 3.x, DeepSeek V3, Qwen 3
- Latency-sensitive single-GPU edge deployments (8 experts × 7B still demands meaningful VRAM)
- Vision / multimodal tasks — text only at this generation