Throughput (tokens / sec)
avg. 210tokens/s
Tool-heavy coding at 3B-active cost
Run instantly. Pay only for what you use.
Qwen 3.6 35B-A3B is Alibaba's sparse open-weight MoE under Apache 2.0: 35B total parameters with just 3B active per token, a 262K-token context window, and native multimodal input — text, images, and video in one model.
It is built for agents that run all day: hybrid Gated DeltaNet attention keeps long contexts cheap, and thinking preservation carries reasoning traces across turns, so multi-step loops keep earlier decisions instead of re-deriving them.
The result is tool-heavy coding at fleet economics: 73.4% on SWE-bench Verified, 80.4 on LiveCodeBench v6, and 81.7 on MMMU with thinking enabled — near-flagship results from 3B-active compute, with strict function calling and JSON output.
3Bactive of 35B total
A 256-expert MoE routes 8 experts per token — near-dense quality at small-model compute.
262Ktoken context window
Holds long agent sessions and full document sets in a single context.
3input modalities
Text, images, and video are read natively — no separate vision model to route to.
Qwen 3.6 35B-A3B is strongest where agents call tools constantly and the context includes more than text — coding loops, visual tickets, and multilingual pipelines at fleet scale.
73.4% on SWE-bench Verified and 51.5% on Terminal-Bench 2.0 with thinking enabled — plans, edits, and verifies real repositories.
81.7 on MMMU and 75.3 on MMMU-Pro — screenshots, diagrams, and video frames are first-class inputs, not an add-on.
Reasoning traces persist across conversation turns, so multi-step agent loops keep earlier decisions without re-deriving them.
92.7 on AIME 2026 and 83.6 on HMMT with thinking enabled — reasoning depth on demand, off when speed matters.
High-volume coding agents and CI automation
Resolves 73.4% of SWE-bench Verified issues with thinking enabled — production coding-agent capability at 3B-active cost.
Multimodal tickets and screen understanding
Screenshots, UI recordings, and diagrams are native inputs — 81.7 on MMMU — so visual bug reports stay in one model.
Tool-heavy loops where per-token cost dominates
3B active parameters and cache-read discounts keep always-on agents and batch pipelines affordable at scale.
Long agent sessions
The 262K window plus thinking preservation holds context and reasoning across hundreds of tool calls.
The hardest cross-app tool orchestration
Scores 37% on MCPMark multi-app orchestration. For marathon runs across many tools, use Kimi K2.6.
Tight output-token budgets with thinking on
Thinking mode is verbose — roughly 4× the median output tokens in Artificial Analysis testing. Disable thinking or cap its budget for cost-sensitive replies.
Audio input or speech understanding
Inputs are text, image, and video only. Transcribe audio upstream before sending it to the model.
Compare Entrim’s pricing, powered by an optimized inference runtime, against other providers using the same token counts per request.
avg. 210tokens/s
avg. 640ms
73.4%SWE-bench Verified
Resolves real GitHub issues end-to-end in agentic coding runs.
80.4LiveCodeBench v6
Competitive coding problems unseen at training time.
81.7MMMU
University-level reasoning over images, charts, and diagrams.
Code examples
curl -X POST "https://api.entrim.ai/v1/chat/completions" \ -H "Content-Type: application/json" \ -H "Authorization: Bearer $ENTRIM_API_KEY" \ -d '{ "model": "Qwen/Qwen3.6-35B-A3B", "messages": [ { "role": "user", "content": "Why do they call it a building if it is already built?" } ] }'SDKs & docs
Use familiar SDK patterns with quickstart examples for common production setups.
OpenAI-compatible
Keep the request format your team already knows.
No cold starts
Production requests are served without model spin-up delays.
Dense agentic coding with thinking control
Gemma 4 quality at 4B-active speed
High-throughput agentic work with 1M context