All news

Cursor Composer 2.5 ships on open-weights Kimi K2.5 at $0.50/M input

Cursor's May 18 Composer 2.5 release keeps Composer 2's Kimi K2.5 backbone and bets RL fine-tuning closes the gap to Opus 4.7 on coding benchmarks. Standard tier is $0.50/M input + $2.50/M output — roughly a tenth of frontier closed-ref pricing. The base weights are public under Modified MIT.

Cursor shipped Composer 2.5 on May 18 2026 — a substantial intelligence/behaviour bump on top of Composer 2, retaining the same Moonshot Kimi K2.5 open-weights backbone. Standard-tier pricing holds at $0.50/M input and $2.50/M output with a faster variant at $3.00/M input and $15.00/M output. Cursor claims parity with Claude Opus 4.7 on SWE-Bench Multilingual (Composer 79.8% vs Opus 80.5%) at roughly a tenth the per-token cost.

The model is K2.5, not K2.6 — and that matters

Cursor’s own announcement is explicit: Composer 2.5 is built on Kimi K2.5, the same checkpoint Composer 2 used. Multiple third-party writeups quote Cursor saying the team kept the K2.5 base and added training rather than swapping to K2.6 (April 2026). The bet is that targeted RL on a known-good base outperforms a base swap for a coding-agent product, where stability and harness compatibility matter as much as raw bench scores.

(K2.6 is the newer Moonshot model — 1T MoE, 32B active, 256K context, native INT4, top of OpenRouter’s coding token volume in May. It’s a strong default for self-hosters, but it’s not what’s powering Composer 2.5.)

What changed on top of K2.5

Cursor names three training shifts:

  • Textual-feedback RL. Instead of only reward-shaping at end-of-run, the trainer surfaces localized hints when tool calls fail mid-task. The model learns to recover inside a session rather than re-rolling.
  • ~25× more synthetic tasks. Including “feature deletion” puzzles — remove a feature, then rebuild it — which forces the model to reconstruct context rather than pattern-match a finished diff. Cursor argues this is closer to the distribution of real engineering work than pure SWE-Bench-style fix tasks.
  • MoE-scale infra rework. Sharded Muon optimizers and dual-mesh HSDP for training-time efficiency. This is the kind of infra change that lets a small team run multiple RL epochs on a 1T-param MoE on a sensible cluster.

None of these change the base K2.5 weights public on HuggingFace; they live in Cursor’s hosted variant. Self-hosters using K2.5 from moonshotai/Kimi-K2.5 won’t get the Composer 2.5 behaviour for free — the RL deltas are Cursor’s IP. The base weights are still the substrate, though, which is the part of the story that matters for the open ecosystem.

Benchmarks (Cursor self-report — treat as vendor numbers)

  • SWE-Bench Multilingual: Composer 2.5 79.8% — vs Opus 4.7 80.5%.
  • CursorBench v3.1 (default settings): Composer 2.5 63.2% — vs Opus 4.7 61.6%.
  • Terminal-Bench 2.0: Composer 2.5 ≈ 69.3% — vs Opus 4.7 ≈ 69.4%.

These are Cursor-published numbers; expect community leaderboards (tbench.ai, benchlm.ai) to land independent scores over the next couple of weeks. Per UPDATE.md discipline, when a vendor’s “Adaptive” benchmark and a community leaderboard diverge by 3+ points, prefer the leaderboard. The Composer-vs-Opus deltas above are small enough that the next independent run could put either model on top — don’t over-anchor on the day-one self-report.

The bigger structural claim is the cost ratio. Opus 4.7 lists at roughly $15/M input and $75/M output on Anthropic’s API; Composer 2.5 standard tier at $0.50/M input and $2.50/M output is ~30× cheaper input and ~30× cheaper output. Even the fast tier ($3/$15) is ~5× cheaper input. If the benchmark parity holds in independent runs, Composer 2.5 lands as the obvious default for high-volume coding-agent workloads where each task is bounded and you can swallow occasional retry cost — which is most of them.

Why this matters for local LLM users

Composer 2.5 isn’t a local-LLM product. But it’s the strongest signal yet that an open-weights base — Modified MIT licensed, available on HuggingFace — can power a production coding-agent UI used by hundreds of thousands of developers. Anthropic, OpenAI, and Google still keep their base weights closed; Moonshot ships them. The Cursor deal validates the open-weights agentic-coding category as deployment-ready, not just research-grade.

Practical caveat: K2.5 at full precision is roughly 1 TB of weights. Quantised to INT4 it’s still ~540 GB — out of reach for the 128 GB GB10 / Strix Halo tier this site mostly covers. Running K2.5 locally needs a multi-node cluster or a 4× DGX Spark / 4× Strix Halo setup. That’s why Cursor (and the OpenRouter usage pattern) routes most K2.5/K2.6 inference through Moonshot’s own API or hosted aggregators rather than local hardware. The open-weights story is “the substrate is public,” not “you can run it on your desk this year.”

Sources