# SWE-1.7 — Cognition's Coding Model

SWE-1.7 is a coding model from Cognition (the company behind Devin, which absorbed Windsurf), announced 2026-07-18. It was RL-trained from a Kimi K2.7 base — an open-weight MIT model that had already been through extensive RL post-training. Cognition's framing is cost-performance: a model that reaches near-frontier coding scores at lower cost than GPT-5.5 or Opus 4.8. It runs only inside Devin (Web, Desktop, CLI), served by Cerebras at 1000 tokens/sec. The weights are not published and there's no API for use in other harnesses.

## What's claimed

Cognition reports three agentic coding benchmarks. On FrontierCode 1.1 Main — Cognition's own benchmark — SWE-1.7 scores 42.3%, just under GPT-5.5's 43.0% and below Opus 4.8's 46.5%, well above the Kimi K2.7 Code base at 30.1%. On Terminal-Bench 2.1 it scores 81.5% versus GPT-5.5's 84.2% and Opus 4.8's 86.9%. On SWE-Bench Multilingual it scores 77.8%, slightly ahead of GPT-5.5's 76.8% but behind Opus 4.8's 84.4%. So the "near GPT-5.5 / Opus intelligence" read holds on the first two benchmarks as *close but under*, and the one benchmark where SWE-1.7 edges past GPT-5.5 is on a multilingual SWE-Bench cut, not on Cognition's own leaderboard where it trails.

The other headline claim is about training method. Because SWE-1.7 came from an already-post-trained Kimi K2.7 and still gained a lot from Cognition's own RL, Cognition argues this challenges the idea of a "post-training ceiling" — that RL can push capabilities further than a base model's prior post-training would suggest.

## How they say they trained it

The post spends most of its length on infrastructure and RL technique rather than the model itself. Four pieces stand out.

Entropy preservation. Long RL runs tend to collapse — the model stops exploring and reward plateaus. Cognition uses top-p sampling to keep low-probability tokens (which sharpen the distribution and kill entropy) out of the optimization target, combined with *sampling distribution replay* so the trainer renormalizes over the same kept-set the rollout sampled from, keeping training-inference divergence bounded. A side effect is that above-threshold tokens get zero gradient, which they say focuses learning on the tokens that carry signal. They also credit the Muon optimizer and removing non-deterministic trainer ops.

Multi-cluster training. RL splits cleanly across clusters because only the trainer needs a single high-bandwidth cluster; inference engines just need current weights. Training ran across four datacenters on three continents (their own GPUs plus providers like Fireworks). Instead of broadcasting full weights, the trainer ships compressed weight deltas through cloud object storage — over 99% smaller per transfer — and inference engines prefetch and apply them in place. Cross-continental updates for a 1T-parameter model reportedly finish in 1-2 minutes with 3-4 seconds of inference pause. Fault tolerance leans on NVIDIA Dynamo for inference rerouting and per-step local checkpointing with peer replication for the trainer.

Data quality. Verifier QA to cut false positives and false negatives, difficulty curation (keep tasks the model solves only occasionally, where the learning signal is), and anti-cheating measures — network-restricted sandboxes stripped of git history, an isolated grading path, exploit-signature checks, and reward 0 for any cheating attempt whether or not it succeeded.

Self-compaction. When a run approaches the context limit, the model summarizes its own working state and resumes from that summary, so rollouts extend past the raw context window — up to six hours in the SWE-1.7 run. An alternating length penalty switches between phases that optimize only for success and phases that penalize going over a token/turn/time budget, so easy tasks compress while hard tasks keep their long-horizon behavior. Cognition first tried self-compaction in its earlier Kevin-32B kernel work.

## Reported behavior

Cognition says SWE-1.7 diverges from its Kimi base in two visible ways: a condensed chain-of-thought (fewer function words, roughly half the words per sentence, attributed to the budget phases) and much heavier codebase exploration before acting — more tool calls, file reads, and searches. In bug-fixes it more often chases root cause and probes edge cases and adversarial inputs, and it settles ambiguous semantics by writing small scripts rather than guessing. The stated trade-off is scope creep: it writes extra tests and touches more files than the task needs, which Cognition frames as an industry-wide pattern as reasoning increases.

## Caveats

The strongest number is on Cognition's own benchmark, and even there SWE-1.7 sits below GPT-5.5 and Opus 4.8. Terminal-Bench 2.1 was run on Cognition's internal evaluation framework (Claude Code for Anthropic models, Codex for OpenAI, Devin CLI for the rest). SWE-Bench Multilingual mixes self-reported competitor numbers with Cognition's own Devin CLI runs. Every result is from the vendor, and the model can't be independently benchmarked outside Devin because there are no public weights and no API.

This is the same [[scaffold-model-fit]] problem in a different direction: the benchmark measures the model *and* the Devin harness it's welded to, not the model alone. Where [[little-coder-scaffold-model-fit]] shows a fixed model's score swinging with the scaffold, SWE-1.7 is trained inside its own scaffold and only shipped inside it, so there's no way to factor the two apart. The general caution about single-number model comparisons applies here too — see [[structured-output-benchmark]] on how one benchmark axis (JSON-Pass vs Value-Accuracy) can flatter a model, and [[incompressible-knowledge-probes]] for an attempt to measure a closed model's capacity without trusting the vendor's own eval.

## Interesting from the discussion

The sharpest point on [HN](https://news.ycombinator.com/item?id=48833866) came from a commenter noting that SWE-1.7's cost-vs-performance chart closely mirrors the one Cursor published for Composer 2.5 — except each chart happens to place its own maker's model best. Both models are RL'd from a Kimi base, and their training data and benchmarks likely come from the same kind of source (Devin/Cursor interaction logs), so both plausibly overfit to their own evals. The commenter's read was overfitting, not deliberate deception. Others were blunter: "build your own benchmark, you can win it," and "benchmarks are just vibes with error bars — wake me up when it survives a week on a real codebase without hallucinating a package that doesn't exist."

Two structural complaints recurred. First, SWE-1.7 is closed-weight and absent from HuggingFace despite descending from MIT-licensed Kimi K2.7, and it's usable only inside the Devin harness with no API — which several people treated as a reason not to bother. Second, Cognition carries reputational baggage from its original 2024 Devin demo being widely called misleading; one commenter who knows people there described real products and smart engineers alongside a habit of avoiding unfavorable comparisons and leaning on in-house benchmarks, and said prior Cognition models underperformed comparable Claude and GLM in their own use. A milder observation: showing competitors' API prices is a strange axis when most users don't pay per-token, and the genuinely notable number might be the Cerebras throughput.