consumer- 12.2B
- 128k
- apache 2.0
- Jul 2024
Scores #
Score per dollar #
3444 pts per $/M input
general_score (62) divided by cheapest input price
($0.02/M).
Higher is better value. [See live pricing](/models/mistral-nemo-12b/pricing).
Run it locally #
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
Can you run it? - reference rigs
| Rig | Q4_K_M | Q8_0 |
|---|---|---|
| 4x H100 80GB (320GB) | fast 1009.4t/s | fast 558.3t/s |
| 4x RTX 5090 (128GB) | fast 539.9t/s | fast 298.6t/s |
| 4x RTX 4090 (96GB) | fast 303.7t/s | fast 168.0t/s |
| 2x RTX 5090 (64GB) | fast 270.0t/s | fast 149.3t/s |
| 2x RTX 3090 (48GB) | fast 141.0t/s | fast 78.0t/s |
| Single RTX 5090 (32GB) | fast 135.0t/s | fast 74.7t/s |
| Mac Studio M4 Ultra 192GB | fast 89.7t/s | fast 49.6t/s |
| Mac Studio M4 Ultra 512GB | fast 89.7t/s | fast 49.6t/s |
| Single RTX 4090 (24GB) | fast 75.9t/s | fast 42.0t/s |
| MacBook Pro M5 Max 128GB | fast 50.5t/s | fast 27.9t/s |
| Single GTX 1080 Ti (11GB) | fast 36.5t/s | offload |
| DGX Spark 128GB unified | fast 20.6t/s | ok 11.4t/s |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | ok 15.4t/s | ok 8.5t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | ok 11.6t/s | slow 6.4t/s |
| NVIDIA Jetson Orin NX 16GB | slow 7.7t/s | slow 4.3t/s |
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
How can a 24GB GPU run a 744B model? It does not load the model into VRAM. The quantized weights (e.g. ~410GB at Q4) sit in system RAM; the GPU only holds the small shared attention and router tensors and accelerates prompt processing. Because GLM 5.2 is a Mixture-of-Experts model, each token activates only ~40B of its 744B params, so llama.cpp streams just those active experts from system RAM to the GPU each token (the -cmoe
offload path).
That makes decode speed bound by system-RAM bandwidth, not GPU bandwidth - single digits on DDR4, which is why these rigs show 3-8 t/s even though they “fit.” A bigger GPU (e.g. 2x 3090) keeps more experts resident on-card and raises tok/s; a smaller GPU still runs it but pays the bandwidth tax. A 744B dense model could not run this way - only MoE’s small-active-params trick makes it possible.
Aggressive quants (1-2 bit) trade accuracy for size - roughly 17% accuracy loss at 2-bit vs full precision, and real long-context work often needs Q5 or Q6 even when lower quants “fit.”
Formula estimates here are conservative; real tuned setups can exceed them (one HN user reports ~6 tok/s on a 512GB DDR4 + 2x 3090 rig).
Download options #
Or run it in the cloud #
Live per-provider pricing, throughput and uptime - refreshed about 1 hour ago via OpenRouter. Click a column to sort.
| Provider | Type | Input $/M | Output $/M | Cache $/M | Tok/s | Latency | Uptime | Value |
|---|---|---|---|---|---|---|---|---|
| DeepInfra | ||||||||
| API | 0.02 | 0.03 | - | - | - | 99.95% | best uptime | |
| Mistral | ||||||||
| API | 0.15 | 0.15 | 0.015 | - | - | 99.94% | ||
| DekaLLM | ||||||||
| API | 0.02 | 0.03 | - | - | - | 98.03% | cheapest | |
| Novita | ||||||||
| avoid | ||||||||
API | 0.04 | 0.17 | - | - | - | 69.74% | Default order: throughput among 95%+ uptime providers, then latency; subscriptions last. Sort by any column. Subscription rows show $/mo in the Value column - per-token columns are "-". Affiliate links are marked sponsored / nofollow. Confirm current pricing on the provider's site before committing.
[Detailed API pricing page + JSON endpoint →](/models/mistral-nemo-12b/pricing)
[See who runs Mistral AI in production →](/adoption/mistral)
Inference cost over time #
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.