Qwen3-Coder: The Most Capable Agentic Coding Model Now Available on Together AI
Together AI ● Covered by 2 sources
Together AI now hosts Qwen3-Coder, a 480B-parameter coding model built for messy, real-world engineering tasks, not just snippets. It reportedly matches Claude Sonnet 4 on tough coding benchmarks while staying open.
Based on reporting by Together AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Qwen3-Coder-480B-A35B-Instruct just landed on Together AI's cloud, and the pitch is pretty specific: this isn't another model that's great at writing a single clean function and useless the moment you ask it to touch three services at once. Together AI is positioning it as a model that can hold an entire codebase in its head — 256K tokens of context natively, stretchable to 1M — and actually work through the kind of tangled, cross-system tasks that make up most professional software engineering.
The numbers back up some of the swagger. On SWE-bench Verified, Qwen3-Coder scored 69.6%, just a hair behind Claude Sonnet 4's 70.4%, and well ahead of the roughly 40-50% range other open models tend to land in. It also edged out Sonnet 4 on agentic browser use (49.9 vs 47.4) and agentic tool use (68.7 vs 65.2), though Sonnet still leads slightly on general agentic coding. That's a notable gap closed for an open-weight model going up against a proprietary frontier system.
Under the hood it's a mixture-of-experts setup: 480 billion total parameters, but only 35 billion active at inference, which is how Together AI can serve something this large without the usual painful latency. It was trained on 7.5 trillion tokens, 70% of it code, using what Together describes as complex reinforcement learning workflows aimed at multi-step reasoning rather than one-shot completions.
The use cases Together highlights are the boring-but-important kind: migrating legacy auth systems across microservices, building features that span backend, frontend, database and deployment pipeline simultaneously, and tracing bugs through distributed systems where the real cause is three services away from the symptom. These are the jobs that usually require a senior engineer's mental model of the whole system, not just pattern-matching on a function signature.
Getting access is deliberately frictionless — it's a standard API call through Together's Python SDK, with a playground, batch API for large refactors, and fine-tuning access if you want to bend it toward your team's own conventions. Together is leaning hard on its infrastructure story here, claiming custom kernels, 4x faster inference, and a 99.9% uptime SLA, essentially arguing that the model quality doesn't matter if you can't actually run 480 billion parameters in production without falling over.
My take — AI-written commentary, not fact-checked reporting
I'll believe the Sonnet 4 comparison when independent evals confirm it, because vendor-published benchmarks from the hosting company selling the API always deserve a raised eyebrow. That said, an open-weight model getting within a point of a top proprietary system on SWE-bench Verified is genuinely the story here, not the marketing copy around it — this is exactly the kind of gap-closing that should worry anyone betting entirely on closed frontier moats for coding tools.
Read more about this at: Together AI
