Skip to content

Model Card|LongCat-2.5

Nov 15, 20261 min
TL;DRLongCat-2.5: 1.6T total parameters, ~48B active per token, 1M context, API pricing input $0.30 / output $1.20 per 1M tokens; DeepSWE v1.1 jumps from predecessor LongCat-2.0's 9.7% to 64.6%, BrowseComp hits 91.9% — the highest in the comparison set; native visual tokens folded into the backbone enable 'visual coding'

Look up & save

Loading dictionary…

🌏 中文版

Model Information

FieldValue
Model IDLongCat-2.5
VendorMeituan / LongCat team
Parameters1.6T total, ~48B active per token (MoE)
Context Window1,000,000 tokens
Input pricing (USD/1M tokens)$0.30 (limited-time price, uncached; cached input $0.006)
Output pricing (USD/1M tokens)$1.20
Open sourceYes (MIT — weights shipped a week after the announcement)
Release date2026-11-01 (announced / API live), 2026-11-09 (weights released)
Official announcementLongCat-2.5 Blog
HuggingFacemeituan-longcat/LongCat-2.5
FamilyLongCat 2.x

Highlights

  • DeepSWE v1.1 jumps from predecessor LongCat-2.0's 9.7% to 64.6% (+54.9 points), Terminal Bench v4.0 goes from 3.0% to 38.5%, and PaperBench goes from 45.2% to 82.1% — the team attributes this to a step change from scaling agentic RL, not incremental gains from more pretraining data
  • BrowseComp hits 91.9%, the highest score in the entire comparison set (which includes Claude Opus 5, GPT-5.6-Sol, and DeepSeek-V4.1-Flash)
  • Visual input is folded directly into the same 1.6T MoE backbone: images are converted to discrete tokens via residual vector quantization (RVQ) and share the same token stream and training objective as text, with no separate vision-encoder adapter — the team calls this "visual coding": reading a screenshot, UI, or chart and writing code in the same reasoning loop
  • W4A4 quantization (4-bit weights and activations) lifts decode throughput 42.77% over a W8A8 baseline, enabling training and inference on domestic AI accelerators with under 80GB of memory per card

Benchmark Results

BenchmarkScorePredecessor (LongCat-2.0)Best competitor
Terminal Bench v2.188.470.8DeepSeek-V4.1-Flash: 90.6
DeepSWE v1.164.69.7DeepSeek-V4.1-Flash: 74.2
PaperBench82.145.2GPT-5.6-Sol: 83.3
HLE (with tools)63.137.5Claude-Opus-5: 64.7
BrowseComp91.979.9DeepSeek-V4.1-Flash: 91.5 (LongCat-2.5 is the highest overall)

⚠️ LongCat-2.5/LongCat-2.0 scores are self-reported by Meituan. Competitor scores come from the official release: some are cited from public technical reports or official leaderboards (marked * in the source), others are run in-house by the LongCat team and are indicative only (unmarked in the source). No independent external reproduction has appeared yet.

Comparison with Predecessor/Competitors

Against LongCat-2.0, the biggest gains cluster in agentic RL tasks: DeepSWE v1.1, Terminal Bench v4.0, and PaperBench all jump several-fold to more than tenfold, not single-digit-percentage tweaks. The team's own explanation is a step change from scaling long-horizon agentic RL across four axes at once — data, environments, algorithms, and harnesses — rather than simply adding more pretraining data.

LongCat-2.5 vs. LongCat-2.0 on three agentic benchmarks
LongCat-2.5 vs. LongCat-2.0 on three agentic benchmarksLongCat-2.5 jumps several-fold to more than tenfold over predecessor LongCat-2.0 on DeepSWE v1.1, Terminal Bench v4.0, and PaperBench Score (%) 50 100 9.7 LongCat-2.0 64.6 LongCat-2.5 3.0 38.5 45.2 82.1 DeepSWE v1.1 Terminal Bench v4.0 PaperBench

Source — LongCat-2.5 Hugging Face model card official evaluation table (2026-11). Scale: baseline x=200, 4 units per point (x = 200 + score × 4), horizontal axis starts at 0.

Against competitors, LongCat-2.5 leads on BrowseComp (91.9%, highest overall) and AutomationBench, but still trails Claude Opus 5 by a meaningful margin on the hardest repo-scale engineering tasks: Terminal Bench v4.0 (38.5 vs. 53.9, a 15.4-point gap) and NL2Repo (60.5 vs. 75.3, nearly 15 points). DeepSeek-V4.1-Flash beats LongCat-2.5 on DeepSWE v1.1 and Terminal Bench v2.1. On pure vision benchmarks, LongCat-2.5 is mid-pack: VideoMME scores 85.2 against Kimi-K3's 90.0. The team's own conclusion: a strong all-round agentic model with a distinctive visual-coding edge, but not yet the top performer on the frontier's hardest engineering tasks.

On pricing, input $0.30 / output $1.20 per 1M tokens (limited-time price), with cached input at just $0.006, puts LongCat-2.5 in a low-to-mid price tier similar to DeepSeek-V4.1-Flash, and clearly cheaper than Claude Opus 5 or GPT-5.6-Sol. The API is compatible with both OpenAI and Anthropic request formats, so it plugs directly into existing harnesses like Claude Code, Codex, and OpenCode — a low-risk entry point for teams who want to try a Chinese frontier model.

Implications for Agent Development

Folding visual understanding into the same token stream means an agent can read a screenshot, a chart, or a rendered result and act on it inside the same reasoning loop that writes code. The team's own t-SNE visualization on a smaller probe model shows visual and text embeddings genuinely interwoven in a shared semantic space, not two systems bolted together.

  • If you're building agents that need to "look at the screen, then write code" (frontend debugging, visual acceptance loops, screenshot-to-code): LongCat-2.5 skips the engineering cost of a separate vision-encoder adapter — you can run the full loop on one model
  • If you're building long-horizon agents whose output needs to be "seen through to completion" (3D scene generation, interactive web pages, data-reconciliation reports): the team's own demos (fluid simulation, 3D showcase, city diorama) show this is a deliberate training direction, not an incidental capability
  • Not a fit: engineering tasks with very large repos or extremely complex multi-step work — LongCat-2.5 still trails Claude Opus 5 by a meaningful margin on Terminal Bench v4.0 and NL2Repo, so Opus 5 remains the steadier choice for now. Also not a fit for teams planning long-term around self-hosted open weights — the announcement explicitly said "no open-weight release has been announced," only to reverse course a week later and ship MIT weights; that commitment itself is not stable

Today's Takeaway

I used to think "folding multimodality into the backbone" was just an engineering choice to skip an adapter. But LongCat-2.5's t-SNE visualization shows visual and text embeddings actually land in the same semantic space — meaning an agent reading a screen and an agent reading a document run on the same reasoning, not two bolted-together systems. The other cognitive diff: open weights aren't a commitment you can take at face value on announcement day. LongCat-2.5 explicitly said "no open-weight release has been announced," then reversed and shipped MIT weights a week later — when tracking open-weight models, the license statement on launch day isn't the final answer.

References