Mistral Large 4 (“Le Chonk”): 1.05T Parameters, API Pricing and the Mac Limit

Rechecked October 10, 2026. Mistral AI has opened Mistral Large 4 as a Public Preview. The official model card describes a granular mixture-of-experts system with 1.05 trillion total parameters, 49 billion active parameters, a 1.6B-parameter vision encoder, and a 1-million-token context window on Mistral’s model profile.¹

The headline number that matters most for Mac users is not 49B. A sparse model can activate only a fraction of its network per token while still requiring access to the full expert pool. At 4 bits per parameter, 1.05T parameters work out to roughly 525GB of raw weight data before quantization metadata, runtime memory, KV cache or macOS are counted.

Large 4 is usable through the preview API today. Public model weights are planned for October 27, 2026 according to Reuters, WSJ and Axios.²³⁴ There is no verified full local Mac setup to recommend yet, and a conventional Q4 representation would already exceed 512GB before overhead.

A real source inconsistency: 49B or 52B active parameters?

The publisher’s current sources contradict each other. Mistral’s versioned model page says 49 billion active parameters; Mistral’s live model alias page and its own October 6 launch post say 52 billion active parameters. Both document a roughly 1.05-trillion-parameter MoE model and a 1.6B vision encoder. As of October 8, Mistral has not explained the discrepancy in those sources; it may be an evolving preview or documentation lag, but neither hypothesis is verified. We therefore show 49–52B as the published range instead of pretending one figure is definitive.

Mac consequence: neither 49B nor 52B active implies a 49B/52B download. A full 1.05T-parameter weight set is roughly 525GB at a theoretical 4-bit floor, before metadata, KV cache, the runtime and the OS. No official downloadable weight files are available today, and Mistral promises release by month-end (October 27 is the reported target, not a guaranteed completed release).

Mistral Large 4 specifications that are actually confirmed

Item	Verified status on Oct. 6, 2026
Model	Mistral Large 4
Version	v26.10
Lifecycle	Public Preview
Architecture	granular mixture of experts, natively multimodal
Total parameters	1.05T
Active parameters	49B (versioned page); 52B (live alias/blog)
Vision encoder	1.6B
Mistral model-card context	1M tokens
OpenRouter deployment context	524,288 tokens
Current Mistral endpoint input	$0.68 / 1M tokens
Cached input	$0.07 / 1M tokens
Output	$2.09 / 1M tokens
Public weights	planned for Oct. 27
Final weight license	not confirmed in the checked sources

The parameter, modality, context and feature claims come from Mistral’s current model card.¹ The 524,288-token number is specifically OpenRouter’s current mistralai/mistral-large-4-0 deployment, not a replacement for Mistral’s own 1M specification.⁵

Why “49B active” does not make Large 4 a 49B local model

MoE architectures route each token through a subset of experts. That is valuable for inference compute: the whole 1.05T network does not have to participate in every token calculation.

Storage is a different problem. Unless a future runtime streams or offloads experts in a particularly aggressive way, the system still needs access to the much larger set of expert weights.

That creates a simple rule for local planning:

• use active parameters as a clue about compute per token;
• use total parameters and checkpoint format as a clue about model storage and memory pressure.

The Mac memory math: 525GB at four bits is only the starting point

These are calculated weight-only estimates, not measured Large 4 RAM usage:

Representation	Calculation	Raw weights
BF16	1.05T × 2 bytes	2.10TB
8-bit	1.05T × 1 byte	1.05TB
6-bit	1.05T × 0.75 byte	787.5GB
5-bit	1.05T × 0.625 byte	656.25GB
4-bit	1.05T × 0.5 byte	525GB
3-bit	1.05T × 0.375 byte	393.75GB
2-bit	1.05T × 0.25 byte	262.5GB

Real quantized files are not just a neat pile of parameter bits. Group scales, metadata, runtime allocations, scratch buffers, the KV cache and the operating system all consume additional memory.

Would a 512GB Mac Studio run it?

A conventional full Q4 build is already larger than 512GB in the weight-only calculation. That makes “512GB should be enough because only 49B parameters are active” a bad assumption.

A theoretical 3-bit representation falls below 512GB, but that is not yet a product recommendation. As of October 7, the available sources still do not establish which quantizations will exist, what quality loss they will have, or whether MLX, llama.cpp or another runtime will efficiently support the final architecture.

For current local models, use the existing Mac unified-memory guide and the Unified-Memory Workbench.⁶⁷

1M context on Mistral, 512K on other deployments

There is no need to force these numbers into a false contradiction.

Mistral’s model card lists 1M tokens.¹ Vercel’s Mistral provider also shows 1M.⁸ OpenRouter currently exposes 524,288 tokens with up to 256K output.⁵ Vals documents a 512K context for the configuration it evaluated.⁹

For production use, the endpoint limit is what matters. If your agent workflow depends on a near-million-token prompt, verify the exact provider rather than relying on the family headline.

Current API price: unusually cheap, but the discount is time-sensitive

Mistral’s Large 4 model page currently displays:¹

• $0.68 / 1M input tokens
• $0.07 / 1M cached input tokens
• $2.09 / 1M output tokens

OpenRouter labels the Mistral endpoint 50% off and shows $1.36 / $0.14 / $4.18 alongside those discounted values.⁵ Treat that as a time-sensitive commercial detail, not a permanent property of the model.

What that means in real token volumes

A workload with 10M input tokens + 1M output tokens costs:

10 × $0.68 + 1 × $2.09 = $8.89

If 80% of the input is billed at the cached-input rate:

2 × $0.68 + 8 × $0.07 + $2.09 = $4.01

At 100M input + 10M output, the same math gives $88.90 without caching or $40.10 with an 80% cache assumption.

These are our calculations, not an estimate of a typical user’s monthly bill. Cache reuse varies dramatically by workload.

EU regional inference is a separate deployment choice

Mistral documents EU and US regional endpoints at a 1.1× price multiplier over standard list pricing.¹⁰ Regional endpoints also have feature differences: the current documentation says Function Calling is the supported regional tool, while stateful features including Agents, Batch and Files are not available there.¹⁰

That makes “EU-hosted” a deployment decision with concrete feature and cost trade-offs, not just a privacy badge.

Try Mistral Large 4 from a Mac: verify the actual API endpoint

Mistral Large 4 is currently available as a hosted Public Preview through Mistral’s Chat Completions API. Running the client on a Mac does not mean local Apple Silicon inference. The model identifier shown in Mistral’s October 6 changelog is mistral-large-4 (do not assume the generic alias mistral-large-latest is pinned to this preview).

Prerequisites: Access to Mistral Studio with billing enabled, curl and jq (install with brew install jq if necessary). API requests may cost money. Use a synthetic prompt only; never paste credentials, private repositories or sensitive client materials into a model smoke test.

# Read the API key without putting it in shell history or a source file.
read -rs MISTRAL_API_KEY
export MISTRAL_API_KEY
# Global endpoint: no guaranteed country of inference processing.
export MISTRAL_BASE_URL=https://api.mistral.ai

# Check that this exact model ID is offered on the selected endpoint.
set -o pipefail
curl --fail-with-body -sS "$MISTRAL_BASE_URL/v1/models" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  | jq -e '.data[] | select(.id == "mistral-large-4") | .id'

# Run a small, synthetic request only after the check succeeds.
curl --fail-with-body -sS "$MISTRAL_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $MISTRAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"mistral-large-4","messages":[{"role":"user","content":"List three criteria for reviewing a code change."}],"max_tokens":100}' \
  | jq '{model, usage, finish_reason: .choices[0].finish_reason}'

When reported by the endpoint, the response includes usage.prompt_tokens, usage.completion_tokens and usage.total_tokens. These are actual fields returned when you run the example, not test data collected by this site. They do not measure Mac inference speed or a model-vs-model win rate. Reconcile invoiced cost with the Mistral usage dashboard/Admin API: aggregated token counts alone cannot prove a cache-read discount.

Select regional inference without assuming country-wide data residency

If you require an EU inference region, explicitly select the EU API hostname and re-run the model-availability check before sending any user content:

export MISTRAL_BASE_URL=https://api.eu.mistral.ai
# Re-run GET /v1/models above first, then POST only if the model is listed.

Mistral’s regional inference policy distinguishes the following:

• Global: api.mistral.ai carries no region-specific inference guarantee. Billing in U.S. dollars or using a U.S. client doesn’t make global processing U.S.-only.
• EU/US inference: api.eu.mistral.ai and api.us.mistral.ai request regional inference for models actually available in that region. If a regional model lookup fails, do not silently fall back to the global host.
• Rates and exclusions: Mistral quotes 1.1× standard list prices for regional requests. Its changelog identifies the Large 4 50%-off launch promotion as lasting two weeks from October 6, 2026. Do not assume promotional prices can simply be multiplied by 1.1; verify the actual contract and regional rate.
• Control plane: Regional inference does not regionalize all data. Account settings, keys, billing and usage analytics can be processed elsewhere. Zero Data Retention (ZDR) is separate. Mistral currently limits regional Agents, Batch and Files features; supported models differ by region.

Production caution: Verify retention, data-processing terms, subprocessors, contract-specific region eligibility and actual logging before submitting regulated data. This article has not executed billable Large 4 API calls or confirmed Large 4’s availability in the EU region.

Large 4 vs Large 3: much more total capacity, modestly more active compute

Mistral Large 3 shipped in December 2025 with 675B total parameters, 41B active parameters, a 256K context window and Apache 2.0 weights.¹¹¹²

	Large 3	Large 4 Preview
Total parameters	675B	1,050B
Active parameters	41B	49B
Official Mistral context	256K	1M
Input price	$0.50/M	$0.68/M current
Output price	$1.50/M	$2.09/M current
Lifecycle	GA	Public Preview
Public weights	yes	planned Oct. 27
License	Apache 2.0	not confirmed yet

Large 4 increases total parameters by roughly 55.6%, while active parameters rise by only about 19.5%. That is the architectural story in one line: Mistral expanded the expert pool much faster than the per-token active footprint.

Do not copy Large 3’s Apache 2.0 label onto Large 4. Frandroid’s launch coverage explicitly lists the Large 4 weight license as unspecified at the time of the announcement.¹³

Benchmark reality: strong launch claims, mixed independent data

Mistral’s launch materials highlight preliminary results on DeepSWE v1.1, Finch/FinWorkBench, Harvey Legal Agent, Dense200 and DIOR-RSVG. Several launch reports reproduce those charts.^13141516

VentureBeat provides an important methodological warning: some competitor scores shown in Mistral’s materials could not yet be independently matched to public benchmark sources for the exact configurations.¹⁶

Independent Vals results

Vals currently reports 48.05% ±1.11 on its aggregate index for the Large 4 Preview, ranking the tested configuration #32 of 44 at the time of checking.⁹

Vals benchmark	Large 4 Preview
Harvey Legal Agent	15.83% ±2.96
Finance Agent v2	54.68% ±0.58
Vibe Code Bench v1.1	78.40% ±3.55
BioMysteryBench	67.04% ±2.67
IOI	45.28% ±3.78
Terminal-Bench 4.0	22.73% ±0.88

Vals lists temperature 1, top-p 0.95, a 256K maximum output and high reasoning effort for the profile, while noting that individual benchmarks can use different providers and parameters.⁹

That evidence supports a more useful conclusion than “Large 4 wins” or “Large 4 disappoints”: performance is workload-dependent, and the model is still a mutable preview.

Public Preview means the model can change underneath you

Mistral’s lifecycle policy says Public Preview models can receive silent updates and have no guaranteed path to General Availability.¹⁷

Axios reports that Mistral plans additional reinforcement learning and safety testing before the October 27 weight release.⁴ Any benchmark, latency observation or behavior note published now therefore needs a date and the word Preview attached to it.

Are the weights downloadable today?

No. As of October 7, the final public Large 4 checkpoint is not available. Reuters states that Mistral Large 4 will be publicly available on October 27; WSJ says the company plans to release the trained weights that day, and Axios describes further work before the release.²³⁴

Without released weights, there is no verified Ollama tag, GGUF download, MLX workflow or final weight license yet.

The useful local questions begin after the checkpoint arrives:

1. What are the official file sizes and formats?
2. What license is attached to the weights?
3. Are FP8, NVFP4 or community GGUF/MLX quantizations available?
4. Which runtimes support the MoE routing correctly?
5. What is actual host/unified-memory overhead?
6. Can experts be streamed or offloaded efficiently?
7. What does real Apple Silicon throughput look like?

Who should care about Large 4 right now?

API-heavy coding and agent workflows: worth testing. Mistral lists Structured Outputs, Function Calling, Document QnA, Agents & Conversations and built-in tools on the model card.¹

Long-context document work: potentially attractive, but verify whether your provider gives you 1M or ~512K context.

EU data-location workloads: Mistral offers EU regional inference, but it carries the 1.1× pricing factor and feature constraints.¹⁰

Local Mac inference: wait for the weights. The model’s storage footprint is the gating factor, not the 49B active count.

The API arrives before a practical Mac deployment

Mistral Large 4 is a useful example of why MoE numbers need context. 49B active parameters describe compute; 1.05T total parameters define a radically different storage problem. At four bits, the weight-only estimate is roughly 525GB.

That makes Large 4 immediately interesting as a hosted model and much less straightforward as a local Mac model. The API is live, the current discounted token rates are aggressive, and Mistral exposes a modern tool/agent stack.¹ For local inference, the meaningful date is October 27: only the released weights can tell us which formats, licenses, quantizations and runtimes are real rather than hypothetical.

For models that can already be downloaded, the Unified-Memory Workbench gives a practical size check. Large 4 needs real file sizes and reproducible Apple Silicon measurements after the weights ship.

Sources and verification limits

The 49B and 52B figures originate from different pages published by Mistral itself. We do not force them into a single unsupported specification. The planned downloadable weights remain a press-reported future release, not a tested local checkpoint. The 525GB Q4 estimate is arithmetic for weights alone, not a Mac benchmark.

Footnotes

1. Mistral – Large 4 versioned model card ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. Reuters – Large 4 launch and planned weights ↩ ↩²

3. Wall Street Journal – Large 4 launch ↩ ↩²

4. Axios – Mistral and open-weight model release ↩ ↩² ↩³

5. OpenRouter – Large 4 provider listing ↩ ↩² ↩³

6. AI on Mac – Unified Memory Guide ↩

7. AI on Mac – Mac RAM calculator ↩

8. Vercel – Large 4 AI Gateway model ↩

9. Vals AI – Large 4 independent evaluation ↩ ↩² ↩³

10. Mistral – regional inference availability and pricing ↩ ↩² ↩³

11. Mistral – Mistral 3 release ↩

12. Mistral – Large 3 model card ↩

13. Frandroid – Large 4 launch report ↩ ↩²

14. The Next Web – Large 4 model overview ↩

15. Numerama – Large 4 report ↩

16. VentureBeat – Large 4 benchmarks ↩ ↩²

17. Mistral – API model lifecycle policy ↩