Models overview

One page per model. Everything on these pages was measured against the live endpoint, so where a model disagrees with its upstream documentation, these pages follow the endpoint.

GET /v1/models is the source of truth for which models exist right now; it currently returns two.

Side by side

	DeepSeek-V4-Flash	Qwen3.8-Flash-Next
Vendor build	DeepSeek-V4-Flash-0731	Qwen3.8-Flash-Next-FP8
Parameters	43 layers, 256+1 experts, 6 activated	125B total / 6B activated, 512 experts, 10+1 activated
Attention	MLA — 64 Q heads, 1 KV head	Hybrid — 36 linear-attention layers + 12 QSA layers
Licence	MIT	Qwen Community License 1.0
Context	1,048,576	262,144
Input	text	text
Streaming	✅	✅
Tool calling	✅	✅
Parallel tool calls	❌	❌
response_format: json_object	✅	✅
response_format: json_schema	❌	❌
Image input	❌	❌
Thinking	✅	✅
Thinks when reasoning_effort omitted	❌	✅
Usable reasoning_effort tiers	minimal low medium high xhigh max	low medium only
usage.reasoning_tokens	✅	❌
system anywhere in messages	✅	❌ first position only
More than one system	✅	❌
developer role	✅	❌

The messages differences trace back to one thing: Qwen ships a Jinja chat template that raises on anything but a single leading system message, and DeepSeek ships no Jinja template at all.

Writing one code path for both

The intersection that works on every model today:

• one system message, at index 0, using the role name system — not developer
• reasoning_effort set explicitly to low or medium, never omitted
• read thinking text from choices[0].message.reasoning
• read thinking token count from usage.completion_tokens_details.reasoning_tokens
• response_format: {"type": "json_object"} for JSON, never json_schema
• text-only content

Common to every model

These hold regardless of which model you call:

Accepted parameters	temperature, max_tokens, top_p, stream, response_format, tools, tool_choice
Dropped silently	stop, seed, logit_bias, logprobs, top_logprobs, top_k, min_p, repetition_penalty
thinking parameter	Rejected with 400 — use reasoning_effort
Tokenizer reported	GPT
Stability	experimental

Anything outside the accepted set is removed before the request reaches the serving backend — no error, no effect. If you need those parameters, run a dedicated endpoint, which passes your body straight through to vLLM or SGLang.