Quick Comparison Table
Verified against the vendors' own model cards and pricing pages in August 2026.
What TSK-1 Found
Neither family has been tested by TSK-1 yet, and neither carries a profile on the hub. The licence and cost analysis on this page is the reliable comparison until they do. See the TSK-1 hub for the families the benchmark currently covers.
The 2026 Reversal: Both Halves of the Old Argument Flipped
For two years the standard framing was simple. Mistral's flagship was the clean Apache 2.0 choice, and Meta's Llama carried a 700-million-monthly-active-user cap. Both halves of that are now wrong, and a team that deploys on the old assumption picks the wrong model.
- On the Mistral side, the licences are attached to different models than most write-ups say. The Apache 2.0 model is Mistral Large 3 (
25.12). Mistral Medium 3.5 (26.04) ships under a Modified MIT licence whose own card describes it as open for commercial and non-commercial use with exceptions for companies with large revenue. Medium 3.5's context is 256K, not the 128K still quoted widely, and it dates to April 2026, not March. - On the Meta side, the frontier line is no longer Llama at all — it is Muse.
llama.comnow redirects to developer.meta.com/ai. Muse Spark 1.2 and Muse Spark 1.1 are API-only with closed weights, served through the Meta Model API. Muse Glimmer is a 30B open model under Apache 2.0. Llama 4 and Llama 3 are now the earlier open-weight generations.
Put those together and the conclusion inverts. Meta's current open model carries a cleaner licence than Mistral's Medium flagship — while Meta's best models moved behind a closed API. The 700M-MAU argument still applies, but only to the older Llama 4 and Llama 3 weights.
TL;DR: Mistral's Apache 2.0 model is Large 3 (675B total / 41B active, 256K context, $0.50/$1.50 per 1M tokens), not Medium 3.5, which is Modified MIT with a large-revenue exception. Meta's open model is now Muse Glimmer, 30B under Apache 2.0, while Muse Spark is API-only with closed weights. Route by task, licence and published cost — and confirm every claim on the vendor's own model card before you deploy.
The Licence Decision Flow, Rebuilt
The old version of this diagram asked one question: will you cross 700M users? That question no longer separates the options, because both vendors now publish an Apache 2.0 model. The question that actually separates them is what you need the weights for.
The two questions that now decide it:
- Do you need downloadable weights? If yes, Meta's strongest tier is off the table — Muse Spark is API-only. Mistral publishes weights across its whole open line.
- Can you live with a revenue clause? If not, the Mistral model to pick is Large 3, not Medium 3.5. Muse Glimmer clears the same bar at a much smaller size.
Licence Comparison
| Licence dimension | Mistral Large 3 | Mistral Medium 3.5 | Muse Glimmer | Llama 4 / Llama 3 |
|---|---|---|---|---|
| Licence name | Apache 2.0 | Modified MIT | Apache 2.0 | Llama Community License |
| Commercial use | ✓ unrestricted | ✓ with a large-revenue exception | ✓ unrestricted | ✓ under 700M MAU |
| User cap | none | none stated | none | 700M monthly active users, parent entity and affiliates |
| Revenue clause | none | ✓ present — read the card | none | none |
| Self-host | ✓ | ✓ | ✓ | ✓ |
| Redistribute fine-tunes | ✓ | ✓ | ✓ | ✓ under the same terms |
| Outputs train competing models | ✓ allowed | check the card | ✓ allowed | restricted |
Read the model card, not the reputation. Every row above changed at least once in the last twelve months, and licences travel with the specific version string — 25.12 and 26.04 are not interchangeable. Mistral's own line spans three different licence regimes at once: Apache 2.0 on Large 3, Modified MIT on Medium 3.5, and closed Premier terms on Codestral (25.08).
The Per-Task Routing Matrix
Nobody deploys "the best model". They deploy a model per job, and the cost consequence of that choice is usually larger than the quality difference. Prices are US dollars per million tokens, input then output, from each vendor's published rate card.
| Task | Reach for | Licence consequence | Cost consequence |
|---|---|---|---|
| Long-context document reasoning | Mistral Large 3, 256K | Apache 2.0, no cap | $0.50 / $1.50 |
| Everyday general work | Mistral Medium 3.5, 256K | Modified MIT, revenue exception | $1.50 / $7.50 |
| Structured reasoning | Magistral Medium, or Magistral Small | Mistral's reasoning line | $2.00 / $5.00, or $0.50 / $1.50 |
| Agentic coding | Devstral 2, or Devstral Small 2 | open coding line; Codestral is closed | $0.40 / $2.00, or $0.10 / $0.30 |
| Meta's coding and agentic tier | Muse Spark 1.2 | closed weights, API only | not published — you cannot model it |
| High-volume classification | Ministral 3 at 3B | open weights, tiny footprint | $0.10 / $0.10 |
| Cheap general throughput | Mistral Small 4, 119B | open weights | $0.15 / $0.60 |
| Small permissive self-host | Muse Glimmer 30B | Apache 2.0, no cap | hardware only |
| Vision plus text | Mistral Large 3 (2.5B vision encoder) | Apache 2.0 | $0.50 / $1.50 |
Three things fall out of that table that a benchmark chart never shows.
The name does not predict the price. Mistral Large 3 costs three times less on input and five times less on output than Mistral Medium 3.5. Large 3 activates 41B of 675B parameters per token, so serving cost tracks the active parameters, not the total. Any budget built on "larger model, larger bill" will be wrong by a factor of five on output.
An unpublished price is a planning cost, not a saving. Muse Spark may well be excellent. It is also the only entry in the table you cannot put in a spreadsheet before you sign, and that matters more than a couple of benchmark points once volume is real.
The cheap rungs are cheap enough to change the architecture. At $0.10 per million tokens in both directions, Ministral 3 at 3B makes it economic to classify, route and pre-filter everything before a frontier model ever sees it — which is usually a bigger cost lever than switching frontier vendors.
Full Mistral Rate Card
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Mistral Large 3 | $0.50 | $1.50 |
| Mistral Medium 3.5 | $1.50 | $7.50 |
| Mistral Small 4 | $0.15 | $0.60 |
| Ministral 3 — 14B / 8B / 3B | $0.20 / $0.15 / $0.10 | same as input |
| Magistral Medium | $2.00 | $5.00 |
| Magistral Small | $0.50 | $1.50 |
| Devstral 2 | $0.40 | $2.00 |
| Devstral Small 2 | $0.10 | $0.30 |
| Codestral | $0.30 | $0.90 |
| Mistral NeMo | $0.15 | $0.15 |
| Mixtral 8x22B (legacy) | $2.00 | $6.00 |
Mistral's product lines beyond the general models: Magistral for reasoning, Devstral 2 and Codestral for coding, Voxtral for audio, OCR 4 for document extraction, and Shieldstral for safety classification.
Meta does not publish a comparable per-token rate card for the Meta Model API on its developer site, so this page carries no Meta price. Anyone quoting one to you is quoting a third party.
Consumer plans, priced separately
Le Chat, Mistral's consumer surface, is billed apart from the API: Free, Pro $14.99/mo, Team $24.99 per user/mo with a $50/mo minimum, Education $5.99 for verified students up to twelve months, and a custom Enterprise tier.
That separation is the quiet cost problem in 2026. A team that wants a chat surface and API access buys both — per lab. Add a second lab for coverage and you are running two consumer subscriptions and two API accounts before anything ships.
Where Taskade Fits: Routing Instead of Betting
Taskade routes across 15+ frontier models from OpenAI, Anthropic, Google, and open-weight providers inside one workspace, with the AI allowance included in the subscription instead of a separate consumer subscription per vendor. Plans are Free, Pro $10/mo billed annually, Business $25, Max $100, and Enterprise $250.
The point is not that any one model is better. It is that this page had to be rewritten because two vendors changed their licence map inside a single year. Three patterns that hold up under that kind of churn:
✓ Pattern 1: route by requirement, not by loyalty. Set the model per agent so a licence change at one lab is a picker change, not a migration. The licence-sensitive agents and the throughput agents do not have to share a vendor.
✓ Pattern 2: pre-filter cheap, reason expensive. A research automation classifies and routes incoming work on a small open model, then hands only the survivors to a frontier model. Because automations are bounded for-each runs over a resolved list, the volume is predictable and so is the bill.
✓ Pattern 3: keep the context in the workspace, not the model. Workspace DNA — Memory feeds Intelligence, Intelligence triggers Execution, Execution creates Memory — means your projects, agents and automations hold the context. Swapping the underlying model does not cost you the accumulated work, and 100+ integrations keep flowing both ways: triggers pull events in, actions push data out.
See 10 Best Open-Source AI LLMs in 2026 for the wider open-weight field and where Mistral, Meta, Qwen, DeepSeek, Kimi and GLM sit alongside each other.
Choose Meta If
Being fair about this matters more than winning the paragraph.
- You want the smallest cleanly licensed open model. Muse Glimmer is 30B under Apache 2.0 with no user cap. That is a genuinely easier legal and hardware story than a 128B or 675B model for anyone shipping on modest infrastructure.
- You are already invested in the Llama fine-tune ecosystem. Llama 4 and Llama 3 remain the most-forked open-weight lineage, and a domain fine-tune that already exists beats a base model you have to adapt. The 700M-MAU cap applies, and for most organisations it never binds.
- You want Meta's newest coding and agentic model and can accept an API. Muse Spark 1.2 is positioned squarely at coding, agentic workflows and multimodal work. If closed weights and unpublished pricing are acceptable in your setup, it is the current Meta answer.
Choose Mistral If
- You need a plain permissive licence at frontier scale. Mistral Large 3 is Apache 2.0 at 675B total and 41B active, with a 256K window and a vision encoder — and it is the cheaper of Mistral's two flagships on both input and output.
- You need a published rate card. Every Mistral model above has a per-token price you can put in a forecast today.
- You want European jurisdiction in the conversation. Mistral is Paris-based, which several regulated buyers weigh — though residency is decided by where inference actually runs, not by the vendor's address.
Final Word: Check the Card, Then Route
The durable lesson from this refresh is not that one lab beat the other. It is that the two most repeated facts about these vendors were both wrong within a year of being written. Apache 2.0 moved from the model people assumed to a different one. Meta's best models stopped being open at all, while its newest open model got a cleaner licence than the incumbent's flagship.
So: read the model card for the exact version string you intend to deploy, price the task rather than the lab, and keep the routing decision somewhere you can change it in an afternoon.
▲ Memory feeds Intelligence. ■ Intelligence triggers Execution. ● Execution creates Memory. Two open-weight lineages. One workspace. The right model for every step.
This is the origin of living software. 🌱
Build with open-weight and frontier models in one workspace →
Related reading
- 10 Best Open-Source AI LLMs in 2026 — Full open-weight field ranking.
- Qwen vs DeepSeek — Chinese open-source frontier duel.
- Kimi vs DeepSeek — Two more open-weight licence stories.
- Kimi vs Claude — Open-weight agentic coding vs frontier chat.
- GPT vs Claude — OpenAI vs Anthropic head-to-head.
- Multi-Model AI Access — How Taskade Genesis routes 15+ models.
- Tools for AI Agents — The built-in agent toolkit.
- TSK-1 hub — The complete model benchmark dataset.