Hardcore Reviews
Hardcore Reviews

LongCat, SOV, LFM: Three Licenses, Three Hardware Routes

Put three open-weight releases on one table and the striking part is not who scores highest. It is that the three licenses look nothing alike. Meituan's LongCat-2.5 ships under a one-page MIT license; SOV-Transcribe-1 from Salesforce's

Published November 17, 20269 min read
<!-- open-weights-license-hardware-comparison-review | review | LongCat, SOV, LFM: Three Licenses, Three Hardware Routes -->

Put three open-weight releases on one table and the striking part is not who scores highest. It is that the three licenses look nothing alike. Meituan's LongCat-2.5 ships under a one-page MIT license; SOV-Transcribe-1 from Salesforce's Missionforce Sovereign AI team ships under the standard Apache-2.0 text; and Liquid AI's LFM2.5-8B-A1B ships under its own LFM Open License v1.0, which ties commercial rights to an annual-revenue threshold: an entity above $10,000,000 in annual revenue is outside the commercial grant. The hardware side splits the same way into three routes: an NPU cluster on domestic accelerators, self-hosted voice infrastructure for enterprises, and consumer edge devices such as laptops and phones. As of November 16, 2026, laying out the model cards, license texts, and API fields of the three Hugging Face repositories yields no performance leaderboard, only Three Licenses and Hardware Routes compared cell by cell. This piece checks only verifiable fields; it makes no selection recommendation, assigns no scores, and runs no benchmarks.

The three coordinates

The repositories are meituan-longcat/LongCat-2.5, salesforce-sovereign-intelligence/SOV-Transcribe-1, and LiquidAI/LFM2.5-8B-A1B. One repository-level detail first: LongCat-2.5 has no official GitHub repository; the weights and the license text live only on Hugging Face. SOV-Transcribe-1 does not sit under the salesforce namespace; requesting the same name there returns 404, and both the repository and the launch post live under the salesforce-sovereign-intelligence team account.

DimensionLongCat-2.5SOV-Transcribe-1LFM2.5-8B-A1B
LicenseMIT (HF tag license:mit, LICENSE text Copyright (c) 2026 Meituan)Apache-2.0 (HF tag license:apache-2.0, model card verbatim License Apache 2.0., no LICENSE file at repo root)LFM Open License v1.0 (HF license: other / lfm1.0, not a standard SPDX entry)
Parameter framingOfficial 1.6T total, about 48B activated per token; safetensors metadata sums to 1,779,752,648,320 (about 1.78T), gap unexplainedModel card 26.6B total, 4.6B active; official blog 26 billion, 4 billion (rounded); HF page 27B paramsREADME 8.3B total, 1.5B active
Modality and positioningNative multimodal MoE, text, image, video input, switchable thinkingNon-streaming ASR, 16 kHz audio, up to 30 minutes (tested limit), seven languagesText-only, reasoning-only, on-device assistant and tool calling
Context or duration1M tokens (product framing)Not a context-window model; handles recordings, not live streams128,000 (extended from 32,768)
Hardware routeDomestic accelerators, NPU platforms; 194-card Atlas A2 sample from a community deployment guideEnterprise self-hosting; official test hardware one RTX PRO 6000 Blackwell (96 GB)Consumer edge; memory footprint below 6 GB (official self-reported)
Release timingHF repo created 2026-11-04, Meituan blog post 2026-11-10Official blog 2026-11-12, technical report marked coming soonMay 28, 2026 (HF createdAt), not a November release

Each cell comes from a tag, a field, or original text. The parameter row deliberately keeps each side's own framing: LongCat-2.5 officially states 1.6T total parameters with about 48B activated per token, while the safetensors metadata sums to 1,779,752,648,320, roughly 1.78T; the two numbers coexist, the gap is officially unexplained, and any technical explanation would be speculation, so this piece writes none. SOV-Transcribe-1 carries three framings side by side: the model card says 26.6B total and 4.6B active, the official blog rounds that to 26 billion total and 4 billion active, and the HF safetensors field says 27B params; this piece labels each source instead of merging them. LFM2.5-8B-A1B's README states 8.3B total and 1.5B active. The release row is equally literal: the LFM2.5-8B-A1B createdAt is May 28, 2026, which makes it a May release, not a November one.

The model bodies differ too. LongCat-2.5 stores visual signals as discrete tokens processed by the same backbone, the mechanism described in the November 10 post. SOV-Transcribe-1 combines a Whisper large-v3 audio encoder, an audio projector, and a Gemma 4 26B A4B MoE text decoder per the model-card framing, with the HF base_model tag pointing at google/gemma-4-26B-A4B-it and MoE routing kept in FP32; its language list is English, French, German, Spanish, Italian, Portuguese, and Dutch. LFM2.5-8B-A1B stacks MoE, GQA, and gated short convolution blocks over 24 layers, 18 double-gated conv plus 6 GQA, and is a reasoning-only model that writes an explicit chain of thought. SOV-Transcribe-1 is also instruction-controlled: it transcribes verbatim by default and adapts formatting to natural-language instructions, and its context biasing accepts keyword glossaries or topic summaries for specialized terms and names without fine-tuning.

The license layer: one page, one standard text, one set of conditions

LongCat-2.5's LICENSE file ships inside the HF repository and is the standard MIT text, with the copyright line Copyright (c) 2026 Meituan and a grant written without restriction; the only condition is that copies and substantial portions keep the copyright notice and the permission notice. The official blog aligns: released under MIT, supporting commercial inference deployment and second-round training without additional authorization. The boundary worth noting is that the MIT text itself contains no trademark or patent terms, and the model card states it does not grant rights to use Meituan trademarks or patents.

SOV-Transcribe-1's licensing rests on two verifiable fields: the HF tag license:apache-2.0 and the model card's verbatim License Apache 2.0. Apache-2.0 is a standard SPDX entry with an explicit patent grant and NOTICE terms. But enumerating the repository root finds no LICENSE file. The tag and the missing file are two facts that this piece keeps side by side: the license is cited as tag plus model card, with the absence of a bundled license text stated as-is.

LFM2.5-8B-A1B's LICENSE file is titled LFM Open License v1.0, runs eleven sections, and states that it is based on the Apache License 2.0. The core is Section 5. Section 1 defines the Threshold as annual revenue of 10 million United States dollars ($10,000,000). Section 5(a) conditions the commercial-use rights on not exceeding the Threshold; Section 5(b) states that commercial use by an entity above the Threshold is not licensed under the agreement; Section 5(c) exempts a Qualified Non-Profit Organization exempt under Section 501(c)(3) for non-commercial and research purposes. Section 2's copyright grant is expressly subject to the Section 5 limitation, and Section 11 terminates the license automatically and immediately on non-compliance, requiring the user to cease all use and delete all copies. The official FAQ adds that the license has no copyleft requirement, so fine-tuned models can stay proprietary. One tension must be kept as-is: the blog's marketing line reads Open-weight - Download, fine-tune, and deploy without restrictions, while the license text sets a revenue gate in Section 5. This piece puts both sentences on the table, treats the license text as authoritative, and does not characterize the license as open source or not; it describes the clauses and leaves the judgment to the reader.

The hardware layer: NPU clusters, self-hosting, consumer edge

LongCat-2.5's official route statement stops at domestic AI accelerators and can be deployed on NPU platforms; the official posts never name the chips used for training or inference, and single-card memory under 80 GB is the constraint the blog names. The 194-card Huawei Ascend Atlas A2 self-hosting sample, split into 2 vision-encoding, 64 prefill, and 128 decode cards in a W4A4 example cluster, comes from the deployment guide on the longcat-2.5-npu branch of the SGLang-FluentLLM repository; that is a community deployment guide, not a Meituan blog post or model-card sentence, and this piece labels it accordingly. Quantization is W4A4 with experts and activations at 4 bit, plus a separate LongCat-2.5-W4A4 repository; the gains over a W8A8 baseline, prefill throughput up 43.3 percent and decode up 42.77 percent, are model-card self-reported. As of the November 16, 2026 capture, the model card offers no GPU deployment guide (media verification read on 2026-11-12); that means the card lacks one, not that Meituan has made any statement about Nvidia routes.

SOV-Transcribe-1's official positioning sentence sets the deployment premise: sovereign, enterprise-ready voice experiences that you retain control over. Two deployment paths exist: a vLLM plugin pinned to vLLM 0.32.0 (vLLM upstream support is still being upstreamed) and Transformers from 5.19 to 5.22 with trust_remote_code=True; long audio defaults to a 600-second cap in the plugin, unlocked to 30 minutes by an environment variable. The official speed test ran on one NVIDIA RTX PRO 6000 Blackwell (96 GB), with 10 to 30 second clips in about 0.44 seconds; a 430M DFlash draft model adds speculative decoding, officially about 2 to 3 times faster with 15 tokens per draft. The model card gives no minimum memory figure, and this piece does not supply one.

LFM2.5-8B-A1B writes its route into its name: A1B stands for about 1.5B active parameters. The README states verbatim a training budget of 38 trillion tokens, a context extended from 32,768 to 128,000, and a vocabulary scaled from 65,536 to 128,000. On the CPU side it self-reports M5 Max decoding at 253 tokens/s, Ryzen AI Max+ 395 at 146 tokens/s, memory below 6 GB, and about 30 tokens/s on phones, with the phrase the fastest model we tested; on the GPU side it self-reports 18.5K output tokens/s at high concurrency and over 1.6B tokens per card per day on one H100 SXM5. All of these are official self-reported figures, reproduced here with that label. The ecosystem row is dense: day-one support for llama.cpp, MLX, vLLM, and SGLang (README self-reported), plus official GGUF, ONNX, MLX, and DSpark format families, DSpark being a 328M draft model at about 2.5 times decode speed with identical output. The README also draws its own boundary: fit for agentic workflows, tool use, and structured outputs; not the best fit for heavy programming or knowledge-intensive question answering without retrieval.

Score framing: in-house, quoted, and incomparable

Each model has one widely repeated score, and none of the three shares a methodology. LongCat-2.5's Terminal Bench v2.1 figure of 88.4 is unstarred, which by the model card's own rule means Meituan's in-house test, described as in-house and indicative only; the Claude Opus 5 entry at 89.1 in the same table is starred, quoted from public reports and leaderboards rather than measured by Meituan. The two numbers are not the same framing, the comparison must carry that caveat, and the pairing must not be written as approaching Opus 5. LongCat-2.5's DeepSWE v1.1 moves from 9.7 on LongCat-2.0 to 64.6, both unstarred and model-card self-reported. SOV-Transcribe-1's 2.28 percent WER comes from the official blog quoting the Artificial Analysis leaderboard with rankings as of November 12, 2026, while the remaining comparison figures come from Salesforce's own runs, with the technical report still marked coming soon; its keyword F1 rises from 71.9 to 96.8 on 15-second earnings clips when only clip keywords are given, officially self-reported. LFM2.5-8B-A1B's fastest in its size class is a README self-reported slogan. Three scores, three harnesses, three methodologies: this piece makes no cross-board comparison and no ranking; who wins is not a question this piece answers.

What can and cannot be checked

The verifiable part is in the open: tags and fields for the three HF repositories, the LICENSE texts of LongCat-2.5 and LFM2.5-8B-A1B, the verbatim model card of SOV-Transcribe-1, and the deployment guide on the longcat-2.5-npu branch of SGLang-FluentLLM can all be re-read by anyone. Popularity fields serve only as availability context: the HF API downloads field reads 1573 for LongCat-2.5, 463 for SOV-Transcribe-1, and 17044 for LFM2.5-8B-A1B (captured via the hf-mirror.com mirror on November 16, 2026); the field definition is unverified, so this piece writes raw values with a timestamp, no cross-model ranking, no increments.

What cannot be checked is left blank: SOV-Transcribe-1's technical report is officially coming soon; LongCat-2.5's training cluster size is unpublished; the 88.4 figure has no known third-party replication; the gap between 1.6T and 1,779,752,648,320 is officially unexplained and this piece does not speculate. Related in-house pieces take different angles: the LongCat-2.5 preview-period article covered the launch narrative, and a separate in-house piece covered LongCat-2.0 on domestic chips; this piece retells neither, and its angle is the cell-by-cell comparison of Three Licenses and Hardware Routes, with each of the three routes read on its own premises.

FAQ

Q1: Do all three licenses allow commercial use? A1: Read them separately. MIT and Apache-2.0 carry no revenue threshold; the LFM Open License v1.0 conditions commercial rights in Section 5(a) on annual revenue not exceeding $10,000,000, states in Section 5(b) that use above the Threshold is not licensed, and exempts 501(c)(3) qualified non-profits for non-commercial and research purposes in Section 5(c).

Q2: Is the LFM Open License v1.0 an open-source license? A2: This piece does not characterize it and describes the clauses only: on the HF side it is license: other with license_name lfm1.0, not a standard SPDX entry; the Section 5 revenue threshold, the Section 11 automatic termination, and the 501(c)(3) exemption are all in the original text; the official FAQ states there is no copyleft requirement. The judgment is left to the reader.

Q3: Why is there no GitHub repository for LongCat-2.5? A3: As of the November 16, 2026 capture, LongCat-2.5 has no official GitHub repository; the weights, the model card, and the LICENSE text all live in the meituan-longcat/LongCat-2.5 repository on Hugging Face.

Q4: Can 88.4 and 89.1 be compared directly? A4: Not as the same framing. 88.4 is Meituan's in-house test, described as indicative only; 89.1 is the starred model-card quotation of Claude Opus 5 from public leaderboards, a different methodology. The pairing requires the caveat and must not be written as approaching Opus 5.

Q5: Does this piece give a selection recommendation or a benchmark ranking? A5: No. The axis is a cell-by-cell comparison of Three Licenses and Hardware Routes; the three scores come from different harnesses and methodologies, so cross-board comparison, selection advice, and rankings are all outside this piece's scope.

Related reading

The local deployment memory planning, the on-device MoE inference, the Kimi K31 Pro weights access review, and the lightweight on-device speech synthesis are adjacent reads with different angles.

This article is AI-assisted and human-edited. Last updated: 2026-11-17

FAQ

Do all three licenses allow commercial use?
Read them separately. MIT and Apache-2.0 carry no revenue threshold; the LFM Open License v1.0 conditions commercial rights in Section 5(a) on annual revenue not exceeding $10,000,000, states in Section 5(b) that use above the Threshold is not licensed, and exempts 501(c)(3) qualified non-profits for non-commercial and research purposes in Section 5(c).
Is the LFM Open License v1.0 an open-source license?
This piece does not characterize it and describes the clauses only: on the HF side it is license: other with license_name lfm1.0, not a standard SPDX entry; the Section 5 revenue threshold, the Section 11 automatic termination, and the 501(c)(3) exemption are all in the original text; the official FAQ states there is no copyleft requirement. The judgment is left to the reader.
Why is there no GitHub repository for LongCat-2.5?
As of the November 16, 2026 capture, LongCat-2.5 has no official GitHub repository; the weights, the model card, and the LICENSE text all live in the meituan-longcat/LongCat-2.5 repository on Hugging Face.
Can 88.4 and 89.1 be compared directly?
Not as the same framing. 88.4 is Meituan's in-house test, described as indicative only; 89.1 is the starred model-card quotation of Claude Opus 5 from public leaderboards, a different methodology. The pairing requires the caveat and must not be written as approaching Opus 5.
Does this piece give a selection recommendation or a benchmark ranking?
No. The axis is a cell-by-cell comparison of Three Licenses and Hardware Routes; the three scores come from different harnesses and methodologies, so cross-board comparison, selection advice, and rankings are all outside this piece's scope.

Related

Hardcore Reviews

TrnRT, Kindi, Baseten: One License Forbids Managed Service

Three inference engine repositories, TrnRT, Kindi and Baseten, on one comparison table, with license terms and hardware bets as the axis rather than a buying recommendation or benchmark ranking. Apache-2.0, Elastic License 2.0 and AGPL-3.0 coexist; one LICENSE says verbatim that the software may not be provided to third parties as a hosted or managed service, and GitHub reports NOASSERTION. Hardware claims from NKI megakernels on Trainium2 to GB300, B300, Quadro K420 and RTX 3080 are labelled as README self-reports. Stars are cumulative totals, with no growth score or single-day delta.

Nov 16, 20269 min read
Hardcore Reviews

qwen-code Preview, OpenBot, free-claude-code: Three Compared

Sets the qwen-code v1.0.0-preview beside OpenBot v1.0.2 and the rolling v6.x of free-claude-code in one table, comparing version states and access boundaries rather than picking a winner. Five rows cover models and providers, where AG-UI and MCP attach, permissions and approvals, stability during Preview, and license points. The free-claude-code column states AGPL-3.0-only, the NOASSERTION value from GitHub, the README disclaimer that it is not affiliated with Anthropic, and that provider terms govern integrations. Star counts are cumulative totals.

Nov 15, 20269 min read