Chat & agents
Qwen3.5 9B
A compact candidate for assistants, extraction and tool-driven workflows.
Input $0.06Output $0.18USD per 1M tokens · Pricing preview
Single A100 candidateModel card ↗
Model evaluation · Private beta
Find your model.
Know your trade-offs.
Build chat, agents and speech with open models. Start with your workload, compare what matters, and choose with confidence.
Founder-reviewed applications. No card required to apply.
Text & agents
Start with a focused shortlist.
Four candidates for common product workloads. Model access and configuration are confirmed during onboarding.
Multilingual workflows
Mistral Small 3.2 24B
Evaluate structured responses, document understanding and multilingual chat.
Input $0.10Output $0.30USD per 1M tokens · Pricing preview
Single A100 · quantizedModel card ↗
Document understanding
Gemma 3 12B
Explore text and image understanding for everyday product features.
Input $0.08Output $0.24USD per 1M tokens · Pricing preview
Single A100 candidateModel card ↗
Reasoning & coding
Qwen3 32B
A larger candidate when complex responses matter more than model size.
Input $0.15Output $0.45USD per 1M tokens · Pricing preview
Single A100 · quantizedModel card ↗
Preview rates assume shared, batched processing and are not live rates or guaranteed quotes. Input tokens are what you send; output tokens are what the model generates. Final pricing depends on volume, context length and capacity validation. Evaluation candidates, not a live availability list. Single-A100 estimates assume 40GB VRAM; quantization, context length and concurrency affect capacity.
Speech
Audio in. Useful output.
Transcription and speech generation are separate workloads, with their own evaluation and pricing.
Speech to text · Pricing preview
Whisper Large v3
Multilingual transcription for recordings, interviews and product workflows.
$0.025 / audio hour
Volume pricing target. Not a live rate or a guaranteed quote.
Pricing & evaluation details
The target assumes a shared service processing 32,000 audio hours per month. Your quote depends on volume, processing requirements and capacity validation.
Speed, word error rate and language coverage on Yamify are not yet benchmarked. Speaker identification and real-time streaming are not included in this target.
Read the model card ↗Text to speech · Evaluation preview
VibeVoice Realtime 0.5B
Microsoft's open-weight model for generating single-speaker speech from streaming text.
English-first. Pricing and latency available after validation.
Read the model card ↗Benchmark status
Evidence before promises.
Published performance results are not available yet. Pricing previews are planning targets, not measured performance.
| What we measure | How we report it | Current status |
|---|---|---|
| Output quality | Task success on a shared test set | Awaiting measurement |
| Response time | Median & p95 time to first token | Awaiting measurement |
| Throughput | Tokens per second at stated concurrency | Awaiting measurement |
| Transcription | Word error rate & processing time by language | Awaiting measurement |
How we will compare models
Every result will identify the exact model revision, precision, context and decoding settings. Text evaluations will use the same prompts at concurrency 1, 4 and 8, reporting errors alongside median and p95 latency. Warm-up requests and cold starts will be reported separately.
Speech evaluations will use reference transcripts and representative recordings for each tested language. We will report word error rate, full processing time and enabled features. Quantized and full-precision models will be evaluated separately.
Explore more open-weight models
Single-A100 candidates
- Qwen3 14B
- Qwen3 30B-A3B (quantized)
- Qwen3.5 27B (quantized)
- Qwen3.5 35B-A3B (quantized)
- Qwen3.6 35B-A3B (quantized)
- Llama 3.1 8B Instruct
- Gemma 3 4B
- Gemma 3 27B (quantized)
- Gemma 4 26B-A4B (quantized)
- Gemma 4 31B (quantized)
- Mistral NeMo 12B
- Phi-4 14B
Multi-A100 candidates
- DeepSeek V3.2
- DeepSeek R1-0528
- DeepSeek V4 Flash
- DeepSeek V4 Pro
- Qwen3.5 397B-A17B
- Qwen3 Coder 480B-A35B
- Qwen3 235B-A22B
- Llama 3.3 70B Instruct
- Kimi K2.6
- Kimi K3
- NVIDIA Nemotron 3 Super 120B-A12B
This is an evaluation catalog, not reserved capacity. Larger models require a separate multi-GPU deployment. GPU count depends on total weights, precision, context and traffic; active MoE parameters alone do not determine memory needs. License and serving-stack compatibility are reviewed before deployment. No simultaneous hosting of this catalog is implied.
Your next step
Bring one real workload.
Find out what fits.
Tell us what you're building, the model you use and the result you need. We'll review your application and help scope an evaluation.