Mistral Large 4 and the 1T parameter frontier

October 6, 2026. Large 4 arrived in public preview. No blog post, no teaser—just an API you can call today. 1T total parameters, 49B active at inference, 38 on the Intelligence Index. That score matches GPT-6 Luna and trails DeepSeek V4.1 Flash by a single point. First time a model built outside the US or China has reached this tier.

What 1T total, 49B active actually means

Most dense models shout a single number: their parameter count. Large 4 gives you two. 1T parameters stored, 49B used for any single forward pass. That isn’t rounding—it’s deliberate.

The trick isn’t new: mixture-of-experts models like Google’s Switch-Coder and DeepMind’s Sparrow already route tokens to small sub-networks. Mistral simplifies this. Train a single dense 1T model, then freeze all but 49B parameters at inference. No routing, no gating—just a static mask learned during training. The active subset still gets optimised for the task, while the rest remain unused.

Memory footprint shrinks from ~2 TB for a dense 1T model to ~98 GB for the active subset. Training cost scales with the active set, so Mistral spent roughly 49× less compute than a dense 1T baseline. Latency stays close to a dense 49B model because the inactive parameters never touch the GPU.

Mistral is betting that sparse scaling can match dense models on benchmarks without matching their training bills.

Multimodal throughput that actually changes what you can build

Previous Mistral models capped at 8 images per request. Large 4 handles 100. That’s a step change.

Context window is 520k tokens, so a 100-image request still leaves ~400k tokens for text. That’s enough for a 200-page PDF plus captions for every image. The model scores 19% on GDP.pdf reasoning, same as MiMo-V2.6-Pro, two points behind Kimi K3. Not groundbreaking, but not a bottleneck either.

Where it matters is cost per task. Processing slide decks or scanned contracts? 100 images in one request means fewer API calls. Mistral’s pricing, $1.36 per 1M input tokens, $4.18 per 1M output tokens, suddenly looks reasonable when you’re not paying for 12 separate calls.

Benchmarking the 38-point club

38 is the current ceiling, shared by GPT-6 Luna and DeepSeek V4.1 Flash. Mistral Large 4 is the third member.

DeepSeek V4.1 Flash holds that one-point lead, but Mistral closed an 18-point gap from Large 3 in a single jump. That’s a leap.

Specialised domains tell a different story. Cyber Index: 50, same as GLM-5.3-Flash, six points behind MiMo-V2.6-Pro. GDP.pdf: 19%, two points behind Kimi K3. Mistral isn’t leading everywhere, but it’s no longer playing catch-up in the broad intelligence tier.

Pricing: the 4× premium and why it might stick

$1.36 per 1M input tokens, $4.18 per 1M output tokens. That’s roughly 4× the cost of open-weights models with similar Intelligence Index scores. Mistral isn’t hiding it: $1.13 per Intelligence Index task, versus $0.25 for open alternatives.

The sparse architecture helps. Lower training cost means Mistral can price closer to marginal cost, but inference still burns memory bandwidth. Every active parameter has to be read from HBM, even if it’s not updated. A dense 49B model would have lower memory traffic: Mistral’s sparse mask adds overhead.

DeepSeek’s dense 1T model has higher training cost, but inference is cheaper per token. Mistral’s pricing reflects a trade-off, lower training cost, higher inference cost.

For batch processing, DeepSeek wins. For multimodal throughput, 100 images in one request, Mistral’s pricing starts to make sense. It’s a bet on specific use cases.

What Mistral Large 4 enables next

Mistral says Large 4 will be the foundation for specialised models. This is a technical roadmap.

Sparse scaling lends itself to distillation. Freeze the 1T model, train a small dense model on its outputs. The active subset can be distilled into a 7B or 13B model without losing too much accuracy.

Fine-tuning is another path. The sparse mask can be re-learned for a new task while keeping most parameters frozen. That’s cheaper than fine-tuning a dense 1T model, and it preserves the multimodal throughput.

The public preview is a signal. Mistral is shipping and iterating. The next model will likely keep the 1T/49B split, but optimise the mask for specific domains.

The state of the 38-point club

Three models: DeepSeek V4.1 Flash, GPT-6 Luna, Mistral Large 4.

| Model | Active params | Architecture | Training cost estimate | |----------------------|---------------|--------------|------------------------| | DeepSeek V4.1 Flash | 1T | Dense | ~$100M | | GPT-6 Luna | 1.2T | Dense | ~$120M | | Mistral Large 4 | 49B | Sparse 1T | ~$20M |

That’s the sparse scaling advantage.

Roadmaps diverge. Mistral’s next move is unclear: will they stick with sparse scaling, or pivot to a hybrid approach?

Why Mistral Large 4 is more than a headline score

Sparse scaling works. Mistral proved it at the 38-point tier. This is a technical alternative to dense 1T+ models.

Multimodal throughput is now a first-class feature. 100 images per request changes what you can build on top of the model. Not every use case needs it, but for document processing, slide decks, or medical imaging, it’s a genuine shift.

Cost-performance trade-offs are real. Mistral’s pricing reflects the sparse architecture: lower training cost, higher inference cost. This is a design choice. For batch processing, dense models win. For multimodal throughput, sparse models have an edge.

What’s still missing from the public preview

No training data disclosure. Mistral hasn’t said what went into the 1T model. Was it filtered web data, synthetic data, or something else? That matters for reproducibility and safety.

No ablation studies on sparsity vs accuracy. How much of the 18-point jump from Large 3 is due to the sparse mask? How much is due to better data or training? Without ablation, it’s hard to separate signal from noise.

No latency benchmarks at scale. Multimodal throughput is great, but if latency spikes at 100 images, it’s not useful for real-time applications.

These gaps shape the real-world evaluation horizon. Mistral shipped first, but the rest of the field will catch up. The next round of benchmarks will separate the sparse scaling winners from the one-hit wonders.

The open question

Mistral Large 4 proves that sparse scaling can compete with dense 1T+ models on raw benchmarks. But can it compete on cost, latency, and fine-tuning flexibility?

That’s the question every enterprise buyer will ask. Mistral’s pricing targets multimodal throughput. If the market values 100 images per request, the 4× premium makes sense. If it doesn’t, Mistral will have to adjust.

Either way, the 38-point club just got more interesting. The race for the next tier is on.

Sources

• Introducing Mistral Large 4
• Mistral Large 4 Preview - Intelligence, Performance & Price ...
• Mistral has released Mistral Large 4, making France home ...