Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › Nanbeige4.2 family

Nanbeige4.2-3B

Hugging Face ↗ Follow changes

Nanbeige Apache 2.0 out 3 months ago

4.17B parameters, text in, text out, 256K tokens of context.

Verified 2026-10-25 Rank #591 in LLMs · score 74.3 method v1 Sources 3 Downloads 30d 117.2k · #300 of 619 tracked

Changed this week: 3 new quantized builds on the Hub (56 in all).2026-10-24

Can I run it?2.5GB at INT4Estimatefits a 16 GB cardmethod
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?117.2kdownloads, 30 days#300 of 619 we track · top 48%+11 likes, downloads −2% this week837 likesHugging Face2026-10-24

Worth reading

Picked by rule from 1 tracked piece: the lab's own post, the most shared, a field report, the latest.

Latest · 3 months ago

Agentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1

Nanbeige4.2-3B and Laguna S 2.1 target agentic workloads at different scales.

  • Nanbeige's default configuration does not enable loop-level KV sharing, potentially requiring separate key/value entries for each pass.
  • Retaining reasoning traces in conversation history risks reaching the model's maximum context length on the very next turn.
kaitchup.substack.com 2026-07-31 2 HN pts 1.4k words

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 56 quantized builds and 5 Spaces.

  • As a service0no API provider listed
  • On own hardware56quantized builds
  • Built on1816 fine-tunes · 2 adapters · 0 merges
  • In apps5Hugging Face Spaces

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 16 GB card at INT4

About 2.5 GB at INT4, 5.0 GB at FP8, 10 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 2.5 GB of weights (estimate; KV cache and activations extra)INT42.5 GBFP8: about 5.0 GB of weights (estimate; KV cache and activations extra)FP85.0 GBBF16: about 10 GB of weights (estimate; KV cache and activations extra)BF1610 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 2.5 GB of weights (estimate; KV cache and activations extra)INT42.5 GBFP8: about 5.0 GB of weights (estimate; KV cache and activations extra)FP85.0 GBBF16: about 10 GB of weights (estimate; KV cache and activations extra)BF1610 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

1 dated article and 29 repo commits so far

Teal is the weights (2026-07-21), amber triangles are 1 dated article (each links to it), grey ticks are repo commits.

AugSepOct2026-09-11: 1 repo commit2026-09-09: 3 repo commits2026-08-26: 1 repo commit2026-08-25: 1 repo commit2026-08-02: 1 repo commit2026-07-28: 1 repo commit2026-07-27: 1 repo commit2026-07-26: 1 repo commit2026-07-25: 1 repo commit2026-07-24: 2 repo commits2026-07-23: 3 repo commits2026-07-22: 3 repo commits2026-07-21: 10 repo commitsAgentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1 · 2026-07-31Weights 2026-07-21 · 2026-07-21Weights 2026-07-21now
AugSepOct2026-09-11: 1 repo commit2026-09-09: 3 repo commits2026-08-26: 1 repo commit2026-08-25: 1 repo commit2026-08-02: 1 repo commit2026-07-28: 1 repo commit2026-07-27: 1 repo commit2026-07-26: 1 repo commit2026-07-25: 1 repo commit2026-07-24: 2 repo commits2026-07-23: 3 repo commits2026-07-22: 3 repo commits2026-07-21: 10 repo commitsAgentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1 · 2026-07-31Weights 2026-07-21 · 2026-07-21Weights 2026-07-21now

What people built on it

74 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

74 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

56 quantized (Hugging Face model tree)16 fine-tunes (Hugging Face model tree)2 adapters (Hugging Face model tree)56 quantized16 fine-tunes2 adapters
56 quantized (Hugging Face model tree)16 fine-tunes (Hugging Face model tree)2 adapters (Hugging Face model tree)56 quantized16 fine-tunes2 adapters
GGUF owao/Nanbeige4.2-3B-GGUF for llama.cpp, Ollama, LM Studio 254.7k downloads MLX MercuriusDream/Nanbeige4.2-3B-mlx-8bit for Apple silicon 3.5k downloads FP8 NullSense/Nanbeige4.2-3B-FP8-Dynamic for vLLM, SGLang 1.4k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

27 rows

Benchmarks 27 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (27)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported AA-LCR · Read long documents58.7—model author2026-10-07
author-reported Agent-IF-Oneday67.5—model author2026-10-07
author-reported Claw-Eval52.2pass^3model author2026-10-07
author-reported Claw-Gym65—model author2026-10-07
author-reported GDPval rubrics74.3—model author2026-10-07
author-reported GPQA Diamond · Reason through hard problems87.4—model author2026-10-07
author-reported HLE w/o Search17.8—model author2026-10-07
author-reported HMMT · Reason through hard problems82.8—model author2026-10-07
author-reported IFBench · Chat and follow instructions54.6—model author2026-10-07
author-reported IMO-AnswerBench · Reason through hard problems67.3—model author2026-10-07
author-reported LiveCodeBench · Write code with agents72.5—model author2026-10-07
author-reported MCP-Atlas57.8—model author2026-10-07
author-reported Office-QA-Pro21.1—model author2026-10-07
author-reported Pinch-Bench-V274.7—model author2026-10-07
author-reported Pinch-Bench-V274.7Daily Tasksmodel author2026-10-07
author-reported SWE-bench Pro · Write code with agents46.9—model author2026-10-07
author-reported SWE-bench Verified · Write code with agents63.6—model author2026-10-07
author-reported SciCode35.6—model author2026-10-07
author-reported Terminal-Bench · Write code with agents44.1—model author2026-10-07
author-reported Terminal-Bench · Write code with agents63.3—model author2026-10-07
author-reported general (claw-eval/Claw-Eval)52.2—Model Card2026-10-24
author-reported diamond (Idavidrein/gpqa)87.4—Model Card2026-10-24
author-reported hle (cais/hle)17.8—Model Card2026-10-24
author-reported MathArena/hmmt_feb_202682.1—Model Card2026-10-24
author-reported SWE_Bench_Pro (ScaleAI/SWE-bench_Pro)46.9—Model Card2026-10-24
Show all 27 rows (2 more)
BenchmarkValueSetting / notesReported byDate
author-reported swe_bench_%_resolved (SWE-bench/SWE-bench_Verified)63.6—Model Card2026-10-24
author-reported terminalbench_2 (harborframework/terminal-bench-2.0)44.1—Model Card2026-10-24

What the card says

1 cited sentence

Summary generated, cited

The model uses a Looped Transformer architecture to increase capacity without adding parameters.[1]

  1. “Looped Transformer architecture reuses the transformer la…”

Generated 2026-10-07 15:13 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

22 fields

Listed in LLMs

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesNanbeigeForCausalLM
Model typenanbeige
Dtypebfloat16
Params by dtypeBF16: 4,169,800,704
Size bandM
MoEno
Context length262,144 (config.json max_position_embeddings)
Languagesen, zh
PapersarXiv:2607.22083
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · NanbeigeForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · NanbeigeForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Nanbeige4.2-3B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)10 GB · Single GPU (<24 GB)
FP8 (1 byte/param)5.0 GB · Single GPU (<24 GB)
INT4 (0.5 bytes/param)2.5 GB · Edge (<4 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

All variants

40 variants

Lineage

Fine-tuned from Nanbeige/Nanbeige4.2-3B-Base (same organization) Family: Nanbeige/Nanbeige4.2 74 derivatives on the Hub

Base model: Nanbeige/Nanbeige4.2-3B-Base

quantized: 56finetune: 16adapter: 2

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
owao/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 owao 60 254.7k
bartowski/Nanbeige_Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 bartowski 39 54.9k
gkraker04/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 gkraker04 0 42.4k
Abiray/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 Abiray 16 14.7k
infosave/Nanbeige4.2-3Bcmf quantized — — — apache-2.0 infosave 1 8.8k
Tdamre/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 Tdamre 1 5.5k
AnkitAI/Parable-Nanbeige4.2-3B-Claude-Fable-5-GGUF quantized GGUF — — apache-2.0 AnkitAI 12 4.9k
NANI-Nithin/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 NANI-Nithin 1 4.7k
Andgihat/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 Andgihat 7 4.1k
mradermacher/Nanbeige4.2-3B-i1-GGUF quantized GGUF — — apache-2.0 mradermacher 2 4.2k
MercuriusDream/Nanbeige4.2-3B-mlx-8bit quantized mlx 4.17B M apache-2.0 MercuriusDream 5 3.5k
iamimmanuelraj/Nanbeige4.2-3B-GGUF quantized GGUF — — other iamimmanuelraj 0 3.2k
mlx-community/Nanbeige4.2-3B-OptiQ-4bit quantized -4bit 4.17B M apache-2.0 mlx-community 3 2.7k
ProCreations/grug-3b finetune — 4.17B M apache-2.0 ProCreations 30 1k
mradermacher/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 mradermacher 0 2.4k
tinyopsec/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 tinyopsec 1 2.1k
litert-community/Nanbeige4.2-3B quantized — — — apache-2.0 litert-community 5 1.8k
Nanbeige/Nanbeige4.2-3B-FP8 quantized -FP8 4.17B M apache-2.0 Nanbeige 10 1.4k
NullSense/Nanbeige4.2-3B-FP8-Dynamic quantized -FP8 4.17B M apache-2.0 NullSense 3 1.4k
nicolasembleton/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 nicolasembleton 1 1.4k
MercuriusDream/Nanbeige4.2-3B-mlx-4bit quantized mlx 4.17B M apache-2.0 MercuriusDream 1 1.3k
webmp3/Sakura-Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 webmp3 1 1.2k
guideboardlabs/Nanbeige4.2-3B-GGUF quantized GGUF — — apache-2.0 guideboardlabs 2 1.1k
WaveCut/Nanbeige4.2-3B-heretic finetune — 4.17B M apache-2.0 WaveCut 8 755
johnhalloran/Nanbeige4.2-3B-mps-fix finetune — 4.17B M apache-2.0 johnhalloran 0 1k
NullSense/Nanbeige4.2-3B-NVFP4-FP8-LoopShield quantized NVFP4 3.71B M apache-2.0 NullSense 2 906
AnkitAI/Parable-Nanbeige4.2-3B-Claude-Fable-5 finetune — 4.17B M apache-2.0 AnkitAI 4 641
Soulfate24/Nanbeige4.2-3B-DSpark-Paretrix quantized GGUF — — apache-2.0 Soulfate24 0 737
MercuriusDream/Nanbeige4.2-3B-mlx-bf16 finetune mlx 4.17B M apache-2.0 MercuriusDream 1 641
Nanbeige/Nanbeige4.2-3B-GPTQ-Int8 quantized GPTQ 1.81B S apache-2.0 Nanbeige 4 435
DevQuasar/Nanbeige.Nanbeige4.2-3B-GGUF quantized GGUF — — — DevQuasar 1 502
OsaurusAI/Nanbeige4.2-3B-JANG_6M finetune — 4.17B M apache-2.0 OsaurusAI 1 480
arunb74/Nanbeige4.2-3B quantized GGUF — — other arunb74 0 510
cudo528/Nanbeige4.2-3B-mlx-bf16 finetune mlx 4.17B M apache-2.0 cudo528 0 470
MercuriusDream/Nanbeige4.2-3B-mlx-6bit quantized mlx 4.17B M apache-2.0 MercuriusDream 1 416
srv-sngh/Nanbeige4.2-3B-mlx-nvfp4 quantized mlx 4.17B M apache-2.0 srv-sngh 0 462
OsaurusAI/Nanbeige4.2-3B-JANG_4M finetune — 4.17B M apache-2.0 OsaurusAI 2 336
LostGentoo/Nanbeige4.2-3B-AWQ-W4A16-ASYM quantized AWQ 4.22B M apache-2.0 LostGentoo 0 421
Irfanuruchi/Nanbeige4.2-3B-CE-v1.0 finetune — 4.17B M apache-2.0 Irfanuruchi 0 403
emperorofrome/Nanbeige4.2-3B-GMCE quantized GGUF 4.17B M apache-2.0 emperorofrome 1 347

Repository changelog

29 commits
changed in the last 7 daysOther: 2026-10-24all changes

Repository changelog 29 commits

DateCommitSHA
2026-09-11 Update modeling_nanbeige.py b82e54b
2026-09-09 Update modeling_nanbeige.py 0e13729
2026-09-09 Update configuration_nanbeige.py 89f3709
2026-09-09 Update modeling_nanbeige.py 021ec9c
2026-08-26 Update README.md 3384e42
2026-08-25 Update README.md f761dfb
2026-08-02 Update README.md 5d54321
2026-07-28 Update README.md f56ec5a
2026-07-27 tokenzier update for rstirp fab06df
2026-07-26 Upload configuration_nanbeige.py f79d0a9
2026-07-25 Upload Nanbeige42_report.pdf e6d5746
2026-07-24 Update README.md a8a131e
2026-07-24 Upload Nanbeige42_report.pdf 5723e48
2026-07-23 Update README.md 460b38d
2026-07-23 Update README.md 87b9ca6
2026-07-23 Update README.md 142572d
2026-07-22 Add community evaluation results for CLAW-EVAL, GPQA, HLE, HMMT_FEB_2026, SWE-BENCH_PRO, SWE-BENCH_VERIFIED, TERMINAL-BENCH-2.0 (#7) 451ed48
2026-07-22 Update README.md f67a83e
2026-07-22 Update README.md 59edd24
2026-07-21 Update README.md 5ff54fb
2026-07-21 Upload modeling_nanbeige.py af90333
2026-07-21 Delete modeling_nanbeige.py 84000d3
2026-07-21 Update README.md 6f27730
2026-07-21 Upload 2 files eeb6288
2026-07-21 Create figures/1 8d101c9
2026-07-21 Update config.json 91ed2c3
2026-07-21 Upload 9 files 5953bfe
2026-07-21 Upload 3 files d5ddbf2
2026-07-21 initial commit 20bbaa2

Articles and coverage

1 piece

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Analysis1

Agentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1
kaitchup.substack.com2026-07-31HN 2 pts · 0 comments1,362 words
Nanbeige4.2-3B and Laguna S 2.1 target agentic workloads at different scales. Nanbeige4.2-3B uses a Looped Transformer with 22 physical decoder layers executed twice to achieve 44 layer executions without storing 44 independent layers, fitting on a 16GB GPU but consuming roughly 44 GiB of KV-cache memory at 256K context. Laguna S 2.1 is a 118-billion-parameter Mixture-of-Experts model activating approximately 8 billion parameters per token, featuring 48 transformer layers and 256 routed experts, requiring vLLM 0.25.0 and a 96 GB GPU to exploit its 256K context length.written, reviewed, cited
  • Nanbeige's default configuration does not enable loop-level KV sharing, potentially requiring separate key/value entries for each pass.
  • Retaining reasoning traces in conversation history risks reaching the model's maximum context length on the very next turn.

Papers

1 paper

Papers 1

arXiv links found in the model card and its tags. "Peer-reviewed" is shown only with evidence: a published venue in OpenAlex (measured) or the authors' own journal reference on arXiv (reported).

Similar models

3 models

Similar models

  • Qwen3-4B · 4.02B
    same size band (M), same task (text-generation), more providers, permissive license
  • Qwen3-4B-Instruct-2507 · 4.02B
    same size band (M), same task (text-generation), more providers, permissive license
  • Qwen3-4B-Thinking-2507 · 4.02B
    same size band (M), same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:41 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25

Follow Nanbeige4.2-3B

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.