models

Qwen3.8 27B

Everything below is a fact about the checkpoint, a measurement with its conditions attached, or a known limit.

in bring-up Step-3.7-Flash (stepfun/step-3.7-flash) — In bring-up. It is listed here when its correctness gates pass on the serving configuration, with its own measured numbers and its own price. No date is promised and nothing is charged for it yet.

Model id: qwen/qwen3.8-27b
Weights: huggingface.co/Qwen/Qwen3.8-27B, Apache-2.0. Qwen trained this model; we serve it.
Serving precision: NVFP4 4-bit weights, with speculative decoding from the checkpoint's own multi-token prediction head. Every build passes the correctness gates before it takes traffic.
Context window: 262,144 tokens, the checkpoint's native window. Input and output share it.
Modality: text + image -> text. Images are sent as image_url content parts and billed as input tokens, at the same $0.38 per million. Output is text. The live roster is authoritative if this page drifts: GET https://api.tiyuvta.ai/v1/models for parameters, and the provider feed at GET /models?schema=openrouter for the input modalities and their prices.
Price: $0.38 input, $0.12 cached input, $2.60 output, per million tokens. Full sheet.
Rate limits: Per account, for abuse control, and raised on request — capacity scales with demand rather than sitting behind a published ceiling. Trial credit runs at a lower rate than purchased credit.

Measurements

sealed 2026-08-15

protocol 984-prompt agentic bench file, greedy, streaming, 128 max output tokens, 3 repetitions per cell, medians reported. No shed requests and no errors across all cells.

Latency and reliability
Measurement	Result	Conditions
Time to first token, cold	0.166s	p50, single request, nothing cached.
Time to first token, light load	0.25s	p50 with four requests in flight. Eight in flight measured 0.40s.
Time to first token, continued conversation	0.13s	Turn two of a 5,690-token conversation, 9 of 9 prefix-cache hits.
Output speed, single stream	132–147 tok/s	Speculative decoding on, build v0.84.1. With speculation off the same configuration measured 69.1 tok/s, which is the floor to plan against. For reference, the fastest published provider on the previous-generation 27B board runs 76 tok/s at 0.47s to first token.
Sustained run	576/576	No errors, no shed requests, -0.27% throughput drift end to end.

These are our own measurements. Run your own during the trial credit and use those numbers instead.

Supported parameters

Available

stream, including stream_options.include_usage
tools and tool_choice
response_format: JSON mode and JSON schema
reasoning_effort, reasoning, include_reasoning
temperature, top_p, seed, stop
logit_bias, logprobs, top_logprobs
max_tokens, billed on tokens generated, not tokens reserved
image_url content parts, billed as input tokens

Read from GET https://api.tiyuvta.ai/v1/models on 2026-08-15. That response is the authority.

Not available

top_k, min_p and the repetition penalties
Embeddings, reranking, audio
Fine-tuning and adapter hosting
Batch and asynchronous job APIs
Server-side conversation storage

The live list is machine-readable: GET https://api.tiyuvta.ai/v1/models returns supported_parameters. That response is authoritative if this page drifts.

Correctness gates

run before every release

Serving NVFP4 4-bit weights is a performance decision, and it is only safe if the output is checked. Each gate below runs against the exact artifact and configuration that serves traffic, and a build that fails one does not ship.

Checkpoint validation: 1,968 tensors checked for shape and layout, with numeric spot checks held inside a 1e-3 relative tolerance.
Kernel coverage: 85 kernel cells exercised green, with no silent fallback path left in the serving configuration.
Prefill and decode agreement: Prefill and decode agree token for token against an eager reference implementation running on the same artifact.
Speculative identity: Speculative decoding output is hash-identical to non-speculative output: 8 of 8 prompts, zero divergences at every draft length tested.
Reference oracle: The vendor's higher-precision release is served side by side as an oracle and the differences are recorded before a build takes traffic.

Known limits

Caching covers continuations only

A cached prefix is reused when a request continues a previous one — the same conversation, one turn later. Fan-out over one shared document with different questions does not hit the cache on this model architecture.

Long cold prompts are prefill-bound

A cold 100,000-token prompt spends seconds reading before the first token. Continuing the same conversation skips that work; fan-out over a shared document does not.

Thinking is on by default

Reasoning tokens are billed as output. Lower reasoning_effort when a task does not need it.

Try it with free credit Read the docs