Meituan LongCat-2.5 Review (2026): 1.6T Params, 1M Context, No Weights

A trillion-and-a-half-parameter model that reads screenshots, writes 3D scenes, costs a sixteenth of what Opus 5.5 costs per token — and publishes every capability claim without a single outside check, while the weights stay a hyperlink to its own blog post. Here is what the documentation actually says, what the vendor's own table actually shows, and what you can verify from outside China today.

By AI Tool Scout · Published November 4, 2026 · Specs checked November 4, 2026

What LongCat-2.5 Actually Is

LongCat-2.5 is Meituan's flagship language and vision model, released on November 2, 2026 — the announcement post carries a November 1 dateline, the API changelog entry is dated November 2, and the company's own X account posted it at 2:35pm on November 2. It replaces LongCat-2.5-Preview, which went live on the LongCat API platform on September 25, 2026, and it builds on LongCat-2.0 from June 2026, the model Meituan described as the first trillion-parameter class system trained end to end on domestic Chinese compute.

The architectural claim worth understanding is not the parameter count. It is that vision is tokenized into the language stream rather than bolted on as a separate encoder. Meituan applies residual vector quantization to native-resolution features from a semantic-and-aligned encoder, producing multi-level discrete visual tokens, and trains the codebook embeddings jointly with the backbone. Images therefore arrive as ordinary token sequences, with two practical consequences the vendor states plainly: visual token IDs can be precomputed and cached offline, and the vision encoder disappears from the training graph, which makes load balancing across data-parallel ranks predictable. The family lineage is documented — the blog cites the LongCat-Next paper for the tokenizer — and the t-SNE validation shown is on a smaller 3B-active proxy model, not on the 1.6T model itself. That is a small honesty worth crediting and a gap worth noting: the shared-embedding-space evidence is scaled down.

What it is not: not open weight today, not a model with a published technical report, not auditable on the training side, and not available outside the LongCat platform and the resellers who proxy it. Meituan's own list of supported harnesses is unusually long for a Chinese release — Claude Code, Codex, Hermes, OpenCode, Cline, Kilo Code, CodeBuddy, Cherry Studio, Chatbox and WorkBuddy all have first-party setup pages — which tells you the intended buyer is a developer running an agent loop, not a consumer typing into a chat box.

The specification, as published

Three rows deserve more than a glance. Context is defined as input plus output, which is why the window and the output ceiling interact: a request that asks for 262,144 output tokens has 786,432 left for input. Most vendors quote input context and output cap separately and leave the accounting to you; Meituan states the sum, and the FAQ confirms it when it says your input (messages + max_tokens) exceeds the model's maximum context window on a context_length_exceeded error. Temperature is not yours to set — out-of-range values for top_p, top_k, repetition_penalty and max_tokens are rejected with a 400, but temperature is silently overridden rather than refused, which matters if you have tuned prompts against a specific value. And W4A4 on both weights and activations is the aggressive end of quantization: 4-bit activations are harder to keep accurate than 4-bit weights, and Meituan claims parity with the unquantized model on all evaluated downstream tasks, which is a claim about its own evaluations.

What changed against LongCat-2.0 and the Preview

The generational jump is real and it is concentrated in agentic coding: Terminal-Bench v2.1 from 70.8 to 88.4, v4.0 from 3.0 to 38.5, AutomationBench from 15.8 to 51.8. Those are the vendor's numbers on the vendor's harness and they are still the largest deltas in the release. The price moved too — $0.30 against LongCat-2.0's $0.75 on input and $1.20 against $2.95 on output, a 60% cut on both legs, which is the opposite direction from most launches.

Two things did not change, and both are worth saying out loud. The parameter count is identical — 1.6T total and ~48B active in both generations, so this is post-training, data and architecture work on the same body rather than a bigger model. And the open-weights precedent broke: LongCat-2.0 is on Hugging Face with an MIT licence and is listed at $0.75 / $2.95, while 2.5 ships as API and web app only.

Pricing: $0.30 In, $0.006 Cached, $1.20 Out

This is the most defensible part of the release, because it is money rather than a benchmark, and because it survives the arithmetic. Meituan's pricing page lists three line items and nothing else — no long-context tier, no vision surcharge, no batch discount, no priority queue.

The worked example: 500M input, 80% cached, 50M output

Same modelled month we use across this site, so the numbers compare directly with our DeepSeek V4.1 Flash and Claude Opus 5.5 reviews.

Three conclusions. The output leg is 65% of the bill — $60 of $92.40 — so at this price ratio the model's verbosity matters more than its input rate, and Meituan publishes no tokens-per-task figure we could check. The cache is worth more than the headline: $117.60 saved against a $210 uncached baseline, and because coding agents resend the whole conversation every turn, the hit rate is the single biggest lever a buyer controls. And the number that should worry procurement is the missing one: every rate on the page is a limited-time discount with no list price and no end date, so today's $92.40 month could become a $210 month without any change to your traffic.

Meituan's own worked example, checked

Meituan publishes one billing example: 600,000 uncached input tokens, 400,000 cached input tokens and 500,000 output tokens. We ran it: 600,000 × $0.30 ÷ 1M = $0.18; 400,000 × $0.006 ÷ 1M = $0.0024; 500,000 × $1.20 ÷ 1M = $0.60. Total $0.7824, which matches the published figure exactly. That is a low bar and a surprising number of vendors fail it — DeepSeek's off-peak card, for comparison, would bill the same request at roughly $0.39. Half a million output tokens for 60 cents is the whole pitch of this model in one line.

The Window, the Output Limit, and the Docs That Moved

Three separate pages of Meituan's documentation, read on November 4, 2026, put the output ceiling at 262,144 tokens: the Model Parameters page calls it the max_tokens limit and the thinking_budget maximum, the Quick Start page says 1M token context window with a maximum output length of 256K tokens, and the FAQ answers LongCat-2.5 outputs up to 256K tokens. Meituan's own Claude Code configuration sets CLAUDE_CODE_MAX_OUTPUT_TOKENS to 262144.

That is not what launch-week coverage recorded. Write-ups published on November 2 described a 128K maximum output, and at least one noted that the Quick Start and FAQ pages then said 128K while the parameter page said 262,144. Two days later all four sources agree on 256K. We cannot tell you which figure was true at 9am on November 2; we can tell you that the documentation changed inside the launch window, that the current answer is 256K, and that anyone quoting 128K is quoting a page Meituan has since edited. Treat a specification that moved within 48 hours as one to re-check before you build a product around it.

The 1,048,576-token context is the more interesting number and the more honest one, because it is defined as input plus output and because the price does not tier. There is no 200K step, no double-rate band above 128K, and no per-image charge listed. For long-document work that is a better structure than most frontier cards, which still reprice the whole request once you cross a threshold.

What a million tokens actually costs here

The caveat is that caching is a platform behaviour, not a contract. Meituan publishes no cache-write rate, no retention window and no statement of how a cache hit is counted, so the 98% discount is only as good as the hit rate you actually observe on your own traffic. Measure it in the billing dashboard before you model a number like $0.89 into a product.

What Meituan's Own Table Shows

The announcement publishes 31 benchmark rows across code agents, general agents and visual tasks, with nine models in the columns: LongCat-2.5 and 2.0, GPT-5.5, GPT-5.6-Sol, Claude Opus 4.8, Claude Opus 5, GLM-5.3, Kimi k3 and DeepSeek V4.1 Flash. The 22 rows below are reproduced from it without adjustment. Read the footnotes before the numbers: LongCat-2.5's own results are three-run averages on a Claude Code harness at a 512K sequence length, and every competitor figure carrying an asterisk is taken from that competitor's own technical report, blog post or leaderboard rather than re-run here.

All values from Meituan's evaluation table at longcat.ai/blog/longcat-2.5, read November 4, 2026. Asterisked competitor scores in the source table are self-reported by those vendors on their own harnesses; unasterisked competitor scores were run by Meituan and are described there as “indicative only”. These are not like-for-like measurements.

Across the 22 rows reproduced above, LongCat-2.5 holds the best score in 3 — BrowseComp at 91.9, 3DCodeBench text-to-3D at 55.4 and ResearchRubrics at 81.0 — leads Claude Opus 5 outright in 8 and trails it in 12; two rows carry no Opus 5 figure at all. Read that as a capability shape rather than a ranking: strong on visual and long-horizon research work, mid-tier on repository-scale engineering.

The asterisk problem

An asterisk in this table means the number came from someone else's paper. That is normal practice and Meituan is more transparent about it than most — the footnote is explicit, and it separately marks which competitor scores it ran itself as indicative only. But it changes what the table can support. When LongCat-2.5 scores 88.4 on Terminal-Bench v2.1 with its own harness, three runs and its own scaffold, and Claude Opus 5 is listed at 89.1 from Anthropic's report with a different scaffold, three-tenths of a point does not tell you the models are equal. It tells you two unrelated measurements landed near each other. The same caution applies in reverse: the 90.6 credited to DeepSeek V4.1 Flash on the same row is DeepSeek's number, not Meituan's measurement of DeepSeek.

This is the single most important thing to understand about the launch. The comparison is not dishonest — it is footnoted — but it is vendor-selected in both directions: Meituan chose which benchmarks to publish, which revisions of them, and which competitor reports to quote. A table where the model lands within 0.7 points of the frontier on the headline row and 15 points behind on a harder revision of the same benchmark family is a table that has been curated, however carefully.

The row Meituan would rather you skipped

Terminal-Bench exists in two revisions on this table and they disagree violently about how close LongCat-2.5 is to the frontier. On v2.1 the model scores 88.4, fourth, behind DeepSeek V4.1 Flash at 90.6, Claude Opus 5 at 89.1 and GPT-5.6-Sol at 88.8, and ahead of kimi-k3 at 88.3 and GPT-5.5 at 83.2. On v4.0 — the harder revision, and the one our Opus 5.5 review treats as the current standard — the same model scores 38.5 against 53.9 for Claude Opus 5. That is a 15.4-point gap, and GLM-5.3 at 41.8 and GPT-5.6-Sol at 37.3 bracket it from both sides.

The generational improvement is genuine either way — v4.0 went from 3.0 to 38.5 — but the launch framing leans on v2.1, where the story is a near-tie, and the press coverage repeated v2.1. If your work looks like the tasks in v4.0, the honest summary is that LongCat-2.5 is a competent mid-tier agent coder, not a frontier one, at a fifth of the price. That may still be exactly what you want. It is a different sentence from the one in the announcement.

DeepSeek at the same price

The comparison Meituan probably regrets including is the last column. DeepSeek V4.1 Flash is listed at the same $0.30 input and $1.20 output at peak — the identical card — and on Meituan's own table it takes Terminal-Bench v2.1 90.6 to 88.4, DeepSWE v1.1 74.2 to 64.6, AutomationBench 54.8 to 51.8, MCP Atlas 84.4 to 83.1, SkillsBench 69.8 to 63.3 and Vision2Web 79.2 to 78.6. DeepSeek also wins on SWE Atlas refactoring (56.7 to 55.2), GDPval-AA rubrics (86.8 to 84.8) and FinSearchComp (64.3 to 63.4), and it is half price off-peak.

LongCat-2.5 wins where the architecture is unusual — 3DCodeBench 55.4 to 53.3, CharXiv 90.2 to 87.7, JobBench 65.4 to 63.2, ResearchRubrics 81.0 to 80.9, LongCatClawBench FORTE 83.3 to 73.5 — and on SpreadsheetBench 2 at 34.9 to 33.0 and PaperBench at 82.1, where DeepSeek has published no figure at all. The pattern is consistent with the design: this is a model for seeing a screenshot and building the thing in it, not for grinding through a 40-file refactor. If you need the second, the vendor's own table sends you to DeepSeek at the same price.

What a Third Party Found: Nothing Yet

We looked for an independent score in three places and found none in any of them.

Artificial Analysis is the reference independent harness on this site — it produced the 58 for Claude Opus 5.5 and the 26 for Mistral Large 4 in our reviews. It has no model page for LongCat-2.5 and none for LongCat-2.5-Preview; both URLs return 404. Its leaderboard lists exactly two LongCat entries, LongCat 2.0 and LongCat Flash Lite. The LongCat 2.0 page is instructive: it records the model as released June 2026, open weights, MIT licence, 1.6T total and 48B active parameters, 1M context — and then prints “Intelligence Updated N/A — Artificial Analysis Intelligence Index Unknown”, with speed at 67.7 output tokens per second and a rank of 43 of 79. The predecessor of this model has never been scored on the index either.

The weights are not out. Hugging Face's public API for the meituan-longcat organisation returned 29 models on November 4, 2026, the newest being LongCat-Flash-Lite-Sparse from August 4; LongCat-2.0, LongCat-2.0-INT8 and LongCat-2.0-FP8 are all there from July 8. There is no LongCat-2.5 and no Preview. GitHub's API for the same organisation returned 38 repositories, the most recently pushed being WBench on October 30 and LongCat-DeepResearch on October 23 — and no LongCat-2.5. On the announcement itself, both the GitHub and HuggingFace badges read “Coming soon” and link back to the blog post they sit on.

No technical report. The blog post is the technical document. It is a substantial one — sections on visual tokenization, asynchronous RL with partial rollouts and gradient-staleness correction, multi-teacher on-policy distillation, supernode-aware parallelism, dual-shard FSDP, train–inference numerical consistency and the W4A4 kernel work — but it cites LongCat-Next, DeepSeek's V4.1 report and several community infrastructure papers rather than publishing a standalone paper of its own. There is no training token count, no cluster size, no data-mix table and no ablation that a reviewer could check.

So the honest position is that LongCat-2.5 has no independent evidence of any kind: no third-party index, no downloadable checkpoint to test, no reproducibility artifacts. That is not evidence it is bad. It is evidence that nobody outside Meituan and its resellers knows, and that the strongest available signal is the vendor's own table — which, as the previous section shows, contains information the launch post did not emphasise.

Where It Wins

Visual coding, which is the whole reason the architecture is interesting. The best number in the entire vendor table, for this model, is 55.4 on 3DCodeBench text-to-3D — ahead of GPT-5.6-Sol at 52.3, Claude Opus 5 at 54.3, GLM-5.3 at 53.4 and DeepSeek at 53.3. It pairs with 78.6 on Vision2Web (behind Opus 5 at 80.2, ahead of GPT-5.6-Sol at 73.5) and 90.2 on CharXiv reasoning with tools. Read together, these say something specific: give this model a screenshot, a mockup or a design and it can produce the code and the scene, which is a job most cheap models cannot do at all because they never had native vision. If your workflow is here is a picture of the dashboard we want, that is a real capability at 30 cents a million input tokens.

Long-horizon research and browsing. 91.9 on BrowseComp — the best figure in the table, above Kimi k3 at 91.2, DeepSeek at 91.5 and Opus 5 at 90.8 — plus 81.0 on ResearchRubrics against 72.9 for Opus 5, 82.1 on PaperBench against 79.1, and 63.1 on Humanity's Last Exam with tools, within 1.6 points of Opus 5's 64.7. The RL pipeline described is explicitly built for long trajectories, with network access disabled during training to stop the model gaming rewards, git history stripped, and an LLM judge auditing the resulting traces. That is the right engineering for the workload, and BrowseComp is the row where it shows.

The price, and the shape of the price. $0.30 / $0.006 / $1.20 with a flat rate at every context length, no image surcharge and a 98% cache discount is a genuinely useful card for the workload it targets. A coding agent is a machine for resending the same repository; at a 2% cache read, the economics of leaving a long conversation open stop being a design constraint. Our modelled month is $92.40, and Meituan's own worked example bills half a million output tokens at $0.60.

Harness compatibility, which is unusual for a Chinese model. Two wire formats on one host, OpenAI and Anthropic, plus first-party setup documentation for Claude Code, Codex, OpenCode, Cline, Kilo Code, Cherry Studio, Chatbox, CodeBuddy, WorkBuddy, CatPaw and Hermes. Pointing Claude Code at this model is an env block, which means a team can evaluate it in an afternoon without writing an adapter. For a model whose main selling point is price, removing the integration cost matters as much as the discount.

Cost of the domestic-compute story, if you are buying for a Chinese deployment. This is the only 1.6T-class model we have seen that publishes W4A4 inference as a first-class engineering result rather than a footnote: +43.3% prefill throughput, +42.77% decode throughput, −15.91% TTFT and −19.72% TPOT against the vendor's own W8A8 INT8 baseline, with a 1.73× average speedup at the GMM kernel. For a buyer in the Chinese cloud ecosystem paying domestic-accelerator rates, that is the relevant number and it is not available from any Western lab.

Where It Loses

The harder coding revision, badly. Terminal-Bench v4.0 at 38.5 against 53.9 for Claude Opus 5 is not a rounding difference; it is a different tier of capability on tasks closer to real repository work. NL2Repo at 60.5 against 75.3 says the same thing from another angle, and SWE Atlas test writing at 47.8 against 62.2 and Codebase QnA at 58.3 against 63.2 confirm it. The two coding rows it wins against Opus 5 are refactoring, 55.2 to 51.9, and PaperBench, 82.1 to 79.1. If your evaluation is "can it take a spec and produce a working repository", the vendor's table says no, not yet.

No weights, against its own family's record. Twenty-nine LongCat repositories sit on Hugging Face, and the previous flagship, LongCat-2.0, carries an MIT licence there with three quantization variants next to it. LongCat-2.5 has a badge that says “Coming soon” and links to a blog post. Meituan may well publish — the 2.0 precedent is strong, and the announcement says releases are planned — but “planned, with no date” is not something a procurement decision can be built on, and anybody who signed a contract on the assumption that this family always opens its weights has been burned by exactly one release.

Every rate is temporary. All three line items are labelled limited-time discounted price and the page adds prices are subject to change. No list price, no end date, no notice period. A 60% cut against LongCat-2.0 is a promotional posture, and the most likely trajectory for a discounted launch price on a model this large is upward. Budget on the assumption that the card moves.

Rate limits you cannot plan against. The documentation publishes a single-request limit (the context window) and the shape of the error — HTTP 429 with retry_after: 60 and a recommendation to use exponential backoff. It publishes no requests-per-minute figure and no tokens-per-minute figure anywhere, and limits are applied per API key. For a batch job that is workable; for a product with a latency SLA it is an unknown that you discover at load.

Access friction for the obvious customer. Accounts in mainland China must complete identity verification or enterprise verification before they can recharge at all, and the tiered volume discount that reaches 40% sits behind the enterprise tier. Outside China you can register by email, so the blocker is commercial rather than geographical — but a Western enterprise procurement review will also note that the inference path runs through Meituan's platform, that no data-residency documentation is published on the pricing or quick-start pages we read, and that the model is exactly the kind of dependency many policies exclude by category.

And the ceiling on everything above: no second opinion. A vendor table with an asterisk convention, a curated benchmark selection and a 15-point spread between two revisions of the same test is a table that needs an outside check before it can carry a purchase decision. As of November 4 there is none. That single fact is why this review lands at 3.5 rather than 4.0, and it will resolve the moment either the weights appear or a harness scores the model.

The Domestic-Chip Claim, Examined

Meituan's infrastructure section states that long-sequence training ran with less than 80 GB of memory per domestic AI accelerator, and that both training and inference were done on domestically produced Chinese chips. It is a specific and technically coherent claim: the parallelism strategy is described as aligning FSDP and expert-parallel degrees to the number of accelerators per supernode, with a context-parallel degree no larger than that, keeping all-gathers and all-to-alls inside the high-bandwidth interconnect. Dual-shard FSDP, hybrid vocabulary parallelism, threshold-triggered dynamic expert replication and Muon optimizer sharding are all described as memory-driven responses to that 80 GB ceiling. The FSDP configuration is reported at 86% higher per-device throughput than a DDP baseline on long-sequence supervised fine-tuning.

None of that is checkable from outside China, and the gaps are specific rather than vague. No accelerator vendor or model is named. No supernode size, no cluster count, no total training tokens, no wall-clock training time. The W4A4 throughput gains are measured against the vendor's own W8A8 INT8 baseline, which is a fair internal comparison and not an external one — it does not tell you how the stack performs against an H-class or TPU-class deployment, only that Meituan's 4-bit path beats Meituan's 8-bit path. And the claim that quantization reaches parity with the unquantized model rests on “all evaluated downstream tasks”, which is the vendor's evaluation set.

What can be said is that the claim is consistent with everything else in the release: an 80 GB-per-chip constraint explains the W4A4 investment, the memory-driven parallelism choices, and the decision to remove the vision encoder from the training graph by pre-tokenizing images offline. It is the sort of constraint that produces this exact set of engineering decisions. That is inference, not verification. Buyers should treat “trained on domestic silicon” as a credible statement of intent from a company with every reason to make it, and not as a measured fact.

What the API Actually Gives You

Operationally this is a cleaner product than the marketing suggests. Two wire formats on one host — api.longcat.chat/openai for the OpenAI SDK and api.longcat.chat/anthropic for the Anthropic one — with the model name LongCat-2.5 in both. Claude Code works through a ~/.claude/settings.json env block that points ANTHROPIC_BASE_URL at the Anthropic endpoint and sets the model, the small-fast model and both the Sonnet and Opus defaults to LongCat-2.5. Codex has a matching guide through the Responses API. Billing is prepaid, only HTTP 200 responses are charged, and the balance does not expire.

The constraints are worth listing because they are the ones that bite in production. Temperature is platform-fixed at 1.0 with thinking on and 0.7 with it off, and other values are overridden rather than rejected — so a prompt tuned at 0.3 will silently run at 1.0. Out-of-range values are 400s for top_p, top_k, repetition_penalty and max_tokens, with a Chinese-language validation message. max_tokens counts output only, while the context window counts input plus output, so the two limits interact. Thinking is toggled per request rather than being always on, with a budget up to 262,144, which is a better design than the always-on adaptive thinking on Claude Opus 5.5. And limits are per API key with no published quota, so the first 429 is how you find the ceiling.

Pros & Cons

5 Best LongCat-2.5 Alternatives

The useful question is not whether LongCat-2.5 is a good model at 30 cents — it plainly is — but whether it is the right one for your loop. These five cover the cases where it is not.

Comparison Table

LongCat-2.5 and LongCat-2.0 rates are from Meituan's and the LongCat API pricing pages, checked November 4, 2026; LongCat-2.0's card is not re-pulled here and its figures come from Artificial Analysis's LongCat 2.0 page, which lists $0.75 input, $2.95 output and a 98% cache discount. The AA Index v5 column is read from Artificial Analysis on November 4 where available — 58 for Opus 5.5, 32 for DeepSeek V4.1 Flash, 26 for Mistral Large 4 — and LongCat 2.5 and 2.0 read not scored because AA lists no index value for either. The TB v2.1 column is Meituan's published table, which is a vendor comparison and not a like-for-like measurement; Kimi's entry is for kimi-k3 rather than K3.1 Pro, and Opus 5.5 does not appear there at all. Competitor token prices are carried from our October and November reviews and were not all re-pulled on November 4 — read “not checked” and “not re-checked” cells as blanks, not figures.

What We Could Not Verify

This release has an unusually long list, because it is a Chinese model with no open artifacts. Ten items, listed rather than smoothed over.

One paragraph on the corporate context, labelled as background rather than anything about the model's quality. Meituan is a Hong Kong-listed delivery and local-services platform, not an AI lab that sells models as its main business, and LongCat is best understood as vertical infrastructure: the company runs agents over merchant catalogues, reviews, logistics and customer service at a scale where a 30-cent input rate is a strategic asset rather than a product margin. That explains two otherwise odd decisions — publishing a 29-row benchmark table aimed squarely at Western agent harnesses, and delaying weights on the one release large enough that open distribution would cost real money to serve. It also explains why the pricing is promotional: Meituan's return on this model comes from what it does inside Meituan, and the public API is partly a recruiting and ecosystem play. None of that makes the model better or worse. It does mean that a roadmap published on a blog post is the only signal you will get about what happens next.

Frequently Asked Questions