English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Cognition Releases SWE-1.7: Kimi K2.7 Base Model Redraws the AI Coding Cost Curve

Forum topic · 小凯 · 2026-07-11

Summary

On July 8, 2026, Cognition (the company behind Devin) released SWE-1.7, an agentic software engineering model built on Moonshot AI's open-source Kimi K2.7 base and refined with large-scale asynchronous reinforcement learning. The model reaches near-frontier intelligence at far lower cost per rollout, sitting on the Pareto frontier of cost vs. performance. Key scores: 42.3% on FrontierCode 1.1 Main (vs. GPT-5.5's 43.0% and Opus 4.8's 46.5%), 81.5% on Terminal-Bench 2.1, and 77.8% on SWE-Bench Multilingual, actually beating GPT-5.5's 76.8%. Technical pillars include entropy-stabilized training with sampling distribution replay, cross-continent training across 4 data centers on 3 continents using weight-delta transfers (99%+ traffic reduction), self-compaction for rollouts up to 6 hours, and strict anti-cheating measures. Cognition argues this refutes the 'post-training ceiling' hypothesis, showing significant RL headroom even on heavily pretrained bases. SWE-1.7 is served at 1000 tokens/sec on Cerebras hardware across all Devin platforms. Risks include Cognition's ownership of its own benchmark, unproven production latency claims, base-model licensing dependence, and infrastructure costs that small teams cannot replicate.

> 2026-07-08/09 · Cognition · AI coding models > Original: https://cognition.com/blog/swe-1-7 > Secondary coverage: https://www.marktechpost.com/2026/07/09/cognition-ai-releases-swe-1-7

TL;DR

On July 8, Cognition (parent company of Devin) released SWE-1.7 — an agentic software engineering model built on the Kimi K2.7 base (an open-source model from Moonshot AI) and post-trained with large-scale asynchronous RL. The fact that the base model itself is a Chinese open-source LLM, and the engineering conclusions that follow, are worth unpacking.

Key numbers: frontier-level performance at much lower cost

SWE-1.7 sits on the Pareto frontier of a cost-vs-performance scatter plot (FrontierCode 1.1 Main score vs. dollars per rollout): near-frontier intelligence at far lower cost than other frontier models.

FrontierCode 1.1 Main (100 high-quality agentic coding tasks)

  • SWE-1.7: 42.3%
  • Opus 4.8: 46.5% (+4.2 pp over SWE-1.7)
  • GPT-5.5: 43.0% (+0.7 pp)
  • Kimi K2.7 Code (base): 30.1% (−12.2 pp)
  • SWE-1.6 (previous gen): 9.4% (SWE-1.7 is a >4x improvement)
  • Terminal-Bench 2.1

  • SWE-1.7: 81.5%
  • Opus 4.8: 86.9% · GPT-5.5: 84.2% · Kimi K2.7 Code: 72.7%
  • SWE-Bench Multilingual

  • SWE-1.7: 77.8% — beats GPT-5.5's 76.8%
  • Opus 4.8: 84.4%
  • Key observation: on multilingual generalization, SWE-1.7 beats GPT-5.5, despite being built on Kimi K2.7.

    Why this RL training worked: four engineering pillars

    1. Entropy-stabilized training

    The biggest risk in vanilla RL is entropy collapse — the rollout distribution narrows until only a few tokens are sampleable and training signal dies. Cognition maintains probability consistency between the rollout and trainer sides: rollouts use top-p sampling and record which tokens were actually retained; the trainer renormalizes using that information to avoid KL divergence explosions.

    A key detail is Sampling Distribution Replay: the retained token sets from rollout are stored and the trainer is forced to redistribute over the same sets, avoiding drift like "this token was 30% at rollout but 5% in the trainer."

    2. Cross-continent multi-cluster training

    Training spanned 4 data centers on 3 continents: one trainer cluster in the US, rollout clusters across three continents, mixing in-house GPUs with third-party inference (e.g., Fireworks).

    Syncing 1T-parameter weights across continents: instead of shipping full weights, Cognition transfers weight deltas only — differences computed every K steps, relayed via cloud object storage (not point-to-point broadcast), cutting transfer volume by 99%+. The inference engine prefetches deltas to CPU memory and pauses only 3–4 seconds to apply them.

    Result: cross-continent weight updates for a 1T-parameter model complete in 1–2 minutes — the infrastructure that makes asynchronous RL feasible at continental scale.

    3. Self-compaction for long tasks

    Training rollouts ran up to 6 hours. The fundamental problem is context exhaustion. Cognition trained the model to generate its own summaries and resume from them, jointly optimizing (1) writing more concise summaries and (2) restoring work state from them.

    This was paired with an Alternating Length Penalty: an Unconstrained phase optimizing only task success, and a Budget phase penalizing solutions exceeding token/turn/tool-call budgets. This compresses solved-task length while preserving long-horizon reasoning on hard tasks.

    4. Data quality and anti-cheating

  • Network-isolated sandboxes (no searching the web for answers)
  • Removing git history and reference artifacts
  • Isolating the scoring path from the agent
  • Programmatic checks for known exploit signatures
  • Any cheating attempt is rewarded 0
  • This is why FrontierCode 1.1 uses blocking criteria — failing solutions score 0, no partial credit.

    Why this matters

    1. The "post-training ceiling" hypothesis is refuted

    Kimi K2.7 had already undergone substantial RL post-training by Moonshot AI before Cognition got it. Yet Cognition still extracted +12.2 pp (FrontierCode) and +8.8 pp (Terminal-Bench). This counters the recent assumption that "frontier models are near the RL ceiling." Cognition's answer: with data, algorithms, and infrastructure all in place, RL still has significant room even on heavily trained bases.

    2. The "open-source base + top-tier RL fine-tuning" model is validated

    The business structure: base from Chinese open source (Kimi K2.7), RL fine-tuning and product by a US company (Cognition), product delivered as SaaS (Devin). If it succeeds commercially, expect more "open-source base + application-layer RL" combos over the next 12 months, rather than only closed "proprietary base + internal RL" models.

    3. Cerebras 1000 TPS is key infrastructure

    SWE-1.7 launched day-one on all Devin platforms (Web/Desktop/CLI) via Cerebras at 1000 tokens/sec. A 1M-token generation task takes ~5.5 hours at 50 TPS on typical GPUs, but only 17 minutes on Cerebras — pushing AI coding from "wait minutes for results" toward real-time streaming.

    4. Contrast with GPT-5.6 Sol's same-week release

    OpenAI's GPT-5.6 series landed July 9, one day after SWE-1.7, on a completely different path:

  • OpenAI GPT-5.6: proprietary base + massive capex + user-education cost (three models to choose from)
  • Cognition SWE-1.7: open-source base + application-layer RL + inference hardware innovation
Neither route is proven "the only right answer," but SWE-1.7 signals that intelligence and cost are independent dimensions that can be optimized separately via model architecture, training method, and inference hardware.

Risks and caveats

1. FrontierCode 1.1 is Cognition's own benchmark. Tasks are public (100 standard, plus a 150-task Extended set), but ownership stays with Cognition. SWE-1.7 does not consistently beat GPT-5.5 elsewhere — Terminal-Bench 2.1 shows it 2.7 pp lower. 2. Cerebras 1000 TPS is a commercial claim, not free. The hardware is expensive, and 1000 TPS may only hold at specific batch sizes. Real-world latency awaits Devin user reports. 3. Base-model dependency risk. If Moonshot AI changes the Kimi K2.7 license or stops maintaining it, Cognition's entire RL pipeline needs rebuilding — an inherent fragility of the open-source-base path. 4. Training cost. Four data centers across continents, 6-hour rollouts, custom MoE optimizer — this infrastructure is not replicable by small teams. SWE-1.7's "low cost" is relative to GPT-5.5's inference pricing, not to a startup's cost of reproducing it from scratch.

Overall, SWE-1.7 is a landmark event rewriting the AI coding cost curve, with a clear answer for the industry: frontier intelligence doesn't require frontier pricing — as long as a team deeply optimizes the RL pipeline, inference hardware, and training infrastructure.

Tags

#cognition#swe-1-7#kimi-k2-7#reinforcement-learning#agentic-coding#devin#cerebras#moonshot-ai

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178346321