The story so far
Weights, every dated article, and repo commits on one rail.
12 repo commits since the weights landed
Teal is the weights (2025-06-03), grey ticks are repo commits.
Models › Qwen3-Reranker family
Qwen3-Reranker-4B
Qwen Apache 2.0 out 17 months ago
4.02B parameters, a query and passages in, relevance scores out, 40K tokens of context. 1 provider serves it from $0.025 per million input tokens.
Verified 2026-10-26 07:27 UTC Rank #29 in Embeddings · score 95.3 method v1 Sources 3 Downloads 30d 2.2M · #60 of 619 tracked
Changed this week: 2 new quantized builds on the Hub (65 in all).2026-10-24
Do people use it?2.2Mdownloads, 30 days#60 of 619 we track · top 10%+5 likes, downloads −9% this week187 likesHugging Face2026-10-26
Where to run it
Live hosted endpoints, cheapest first, USD per million tokens. Ties break on output price, then provider name.
| # | Cheapest live row | In /1M | Out /1M | Context | Quant | Seen |
|---|---|---|---|---|---|---|
| 1 | DeepInfra | $0.025 | — | 32K | — | Oct 26 |
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.
Served by 1 API provider; heavily run locally (65 quantized builds).
- As a service1DeepInfra
- On own hardware65quantized builds
- Built on1814 fine-tunes · 4 adapters · 0 merges
- In apps19Hugging Face Spaces
Top Spaces running it: embedl/hfviewer ♥ 181 · kaengreg/rusBEIR ♥ 6 · Agents-MCP-Hackathon/mcp-data-analyst ♥ 5 · emanforhelp/model-pulse ♥ 5
Will it fit?
Weights only, an estimate: leave room for the KV cache. Method
Fits a 16 GB card at INT4
About 2.4 GB at INT4, 4.8 GB at FP8, 9.7 GB at BF16: weights only, so leave room for the KV cache.
Estimate params × bytes per param × 1.2 · how we estimate
What people built on it
83 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.
83 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.
Reference
Everything else, folded: the full tables, every source.
What the card says
1 cited sentence
Use-case evidence
8 use cases
What it is good at 8 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.
- Rerank search resultsstructuraldecision-model estimate p=0.99name: Reranktag: text-ranking+1 more
- Run on a laptop or phonestructuraldecision-model estimate p=0.83name: 4Bfits: single-gpu
- Call tools and APIsstructuralchat template: <tool_call>
- Reason through hard problemsstructuralchat template: <think>
- Chat and follow instructionsstructuralchat template present
- Embed text for searchstructuraltag: sentence-transformers
- Classify and make typed decisionsstructuraldecision-model estimate p=0.88
- Work across languagesstructuraldecision-model estimate p=1.00
Facts
21 fields
Listed in Embeddings
Facts
- Qwen3-Reranker-4B has 4.02B parameters. (safetensors metadata, observed 2026-10-26)
- Context length is 40,960 tokens (40K). (config.json max_position_embeddings, observed 2026-10-26)
- License: Apache 2.0 (friction: none). Permissive; patent grant. (license file, observed 2026-10-26)
- Created 2025-06-03, last modified 2026-04-16. (HF API, observed 2026-10-26)
| Pipeline tag | text-ranking |
|---|---|
| Library | transformers |
| Architectures | Qwen3ForCausalLM |
| Model type | qwen3 |
| Dtype | bfloat16 |
| Params by dtype | BF16: 4,021,784,576 |
| Size band | M |
| MoE | no |
| Context length | 40,960 (config.json max_position_embeddings) |
| Papers | arXiv:2506.05176 |
| License | Apache 2.0 (apache-2.0) |
| Hub snapshot | 2026-10-26 |
Servability
vLLM · SGLang
Servability
vLLM: supported · Qwen3ForCausalLM · 4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60SGLang: supported · Qwen3ForCausalLM · 60d941293e3af27a12c6eb64e2c3e2e1a17f100f
vLLM registry (4b1c04e76fe9ac3ca8b7e541dd423e7d5f38dc60) · SGLang registry (60d941293e3af27a12c6eb64e2c3e2e1a17f100f) · 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 Qwen3-Reranker-4B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.
| BF16 (2 bytes/param) | 9.7 GB · Single GPU (<24 GB) |
|---|---|
| FP8 (1 byte/param) | 4.8 GB · Single GPU (<24 GB) |
| INT4 (0.5 bytes/param) | 2.4 GB · Edge (<4 GB) |
Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.
All variants
40 variants
Lineage
Base model: Qwen/Qwen3-4B-Base
quantized: 65finetune: 14adapter: 4
Top variants (40 shown, ranked by downloads)
| Variant | Relation | Quant | Params | Band | License | Author | Likes | Downloads |
|---|---|---|---|---|---|---|---|---|
| boboliu/Qwen3-Reranker-4B-W4A16-G128 | quantized | -W4A16 | 4.05B | M | apache-2.0 | boboliu | 2 | 2.9M |
| danielchalef/Qwen3-Reranker-4B-seq-cls-vllm-fixed | finetune | — | 4.02B | M | apache-2.0 | danielchalef | 0 | 81.1k |
| dolfsai/Qwen3-Reranker-4B-seq-cls-vllm-W4A16 | quantized | -W4A16 | 4.05B | M | apache-2.0 | dolfsai | 1 | 71.6k |
| Voodisss/Qwen3-Reranker-4B-GGUF-llama_cpp | quantized | GGUF | — | — | apache-2.0 | Voodisss | 18 | 36.2k |
| DevQuasar/Qwen.Qwen3-Reranker-4B-GGUF | quantized | GGUF | — | — | — | DevQuasar | 9 | 28.5k |
| drawais/Qwen3-Reranker-4B-AWQ-INT4 | quantized | AWQ | 4.05B | M | apache-2.0 | drawais | 3 | 21.3k |
| mradermacher/Qwen3-Reranker-4B-GGUF | quantized | GGUF | — | — | apache-2.0 | mradermacher | 9 | 20.6k |
| dolfsai/Qwen3-Reranker-4B-seq-cls-vllm-W4A16_ASYM | quantized | -W4A16 | 4.08B | M | apache-2.0 | dolfsai | 1 | 9.3k |
| furiosa-ai/Qwen3-Reranker-4B | finetune | — | 4.02B | M | apache-2.0 | furiosa-ai | 0 | 8.7k |
| giladgd/Qwen3-Reranker-4B-GGUF | quantized | GGUF | — | — | apache-2.0 | giladgd | 0 | 6.6k |
| DCC-BS/Qwen3-Reranker-4B-FP8-Dynamic | quantized | -FP8 | 4.41B | M | — | DCC-BS | 0 | 6.3k |
| hq-bench/coreb-code-reranker | adapter | — | 4.02B | M | apache-2.0 | hq-bench | 7 | 5.4k |
| IAAR-Shanghai/MemReranker-4B | finetune | — | 4.41B | M | apache-2.0 | IAAR-Shanghai | 8 | 3.3k |
| vserifsaglam/Qwen3-Reranker-4B-4bit-MLX | quantized | -4bit | 4.02B | M | apache-2.0 | vserifsaglam | 2 | 2.6k |
| prithivMLmods/Qwen3-Reranker-4B-F32-GGUF | quantized | GGUF | — | — | apache-2.0 | prithivMLmods | 1 | 2.3k |
| tensorblock/Qwen_Qwen3-Reranker-4B-GGUF | quantized | GGUF | — | — | apache-2.0 | tensorblock | 0 | 2.1k |
| mradermacher/Qwen3-Reranker-4B-i1-GGUF | quantized | GGUF | — | — | apache-2.0 | mradermacher | 0 | 1.9k |
| tatsuyaaaaaaa/Qwen3-Reranker-4B-gguf | quantized | gguf | — | — | apache-2.0 | tatsuyaaaaaaa | 0 | 1.8k |
| dean2155/Qwen3-Reranker-4B-Q8_0-GGUF | quantized | Q8_0 | — | — | apache-2.0 | dean2155 | 1 | 1.4k |
| gscoppino/Qwen3-Reranker-4B-GGUF-llama_cpp | quantized | GGUF | — | — | apache-2.0 | gscoppino | 0 | 1.1k |
| zenlm/zen-reranker-4B-GGUF | quantized | GGUF | — | — | apache-2.0 | zenlm | 0 | 1k |
| batiai/Qwen3-Reranker-4B-GGUF | quantized | GGUF | — | — | apache-2.0 | batiai | 1 | 924 |
| Hanno-Labs/bosun-4b-GGUF | quantized | GGUF | — | — | apache-2.0 | Hanno-Labs | 0 | 786 |
| gorse-io/Qwen3-Reranker-4B-GGUF | quantized | GGUF | — | — | apache-2.0 | gorse-io | 0 | 752 |
| KoliaNik/Qwen3-Reranker-4B-NVFP4 | quantized | NVFP4 | 4.02B | M | apache-2.0 | KoliaNik | 2 | 539 |
| sinjab/Qwen3-Reranker-4B-F16-GGUF | quantized | GGUF | — | — | apache-2.0 | sinjab | 0 | 595 |
| ebasKing/Qwen3-Reranker-4B-GGUF-llama_cpp | quantized | GGUF | — | — | apache-2.0 | ebasKing | 0 | 479 |
| Lipdog/Qwen3-Reranker-4B-mlx-fp16 | finetune | mlx | 4.02B | M | apache-2.0 | Lipdog | 0 | 472 |
| soichisumi/qwen3-reranker-4b-mlx-4bit | quantized | mlx | 4.02B | M | apache-2.0 | soichisumi | 0 | 450 |
| itsbohara/qwen3_reranker_4b_q8_0 | quantized | GGUF | — | — | apache-2.0 | itsbohara | 0 | 442 |
| shawnw3i/Qwen3-Reranker-4B-ONNX | quantized | — | — | — | apache-2.0 | shawnw3i | 0 | 422 |
| novelaide/Qwen3-Reranker-4B-Q4_K_M-GGUF | quantized | GGUF | — | — | apache-2.0 | novelaide | 0 | 421 |
| ebasKing/Qwen3-Reranker-4B-GGUF | quantized | GGUF | — | — | apache-2.0 | ebasKing | 0 | 407 |
| smitxxiv/Qwen3-Re4B-SQL-TeCoD-TMM | finetune | — | 4.02B | M | apache-2.0 | smitxxiv | 0 | 384 |
| enacimie/Qwen3-Reranker-4B-Q8_0-GGUF | quantized | Q8_0 | — | — | apache-2.0 | enacimie | 0 | 368 |
| sinjab/Qwen3-Reranker-4B-Q4_K_M-GGUF | quantized | GGUF | — | — | apache-2.0 | sinjab | 0 | 368 |
| Hanno-Labs/bosun-4b | adapter | — | — | — | apache-2.0 | Hanno-Labs | 1 | 311 |
| ExpedientFalcon/Qwen3-Reranker-4B-Q4_K_M-GGUF | quantized | GGUF | — | — | apache-2.0 | ExpedientFalcon | 0 | 293 |
| zenlm/zen-reranker | finetune | — | 4.02B | M | apache-2.0 | zenlm | 1 | 237 |
| MaYiding/EventTwin | finetune | — | 4.02B | M | apache-2.0 | MaYiding | 2 | 166 |
Repository changelog
12 commits
Repository changelog 12 commits
| Date | Commit | SHA |
|---|---|---|
| 2026-04-16 | Integrate with Sentence Transformers v5.4 (#11) | 22e6836 |
| 2025-06-09 | update model type and add paper link (#3) | f16fc5d |
| 2025-06-06 | Update README.md | 5790622 |
| 2025-06-05 | Update README.md | 80be864 |
| 2025-06-05 | Update README.md | c12eecf |
| 2025-06-05 | Update README.md | 6896691 |
| 2025-06-05 | Delete added_tokens.json | 501a274 |
| 2025-06-05 | Update README.md | 5c76dc8 |
| 2025-06-05 | Update README.md | ff536d3 |
| 2025-06-05 | Update README.md | 74c270b |
| 2025-06-03 | Upload folder using huggingface_hub | ccd0ad1 |
| 2025-06-03 | initial commit | fcc16dd |
All providers
1 provider · 1 endpoint
Where to run 1 providers · 1 endpoints
Hosted endpoints observed for Qwen3-Reranker-4B, sorted by input price (USD per million tokens). Each row names its observer and date; prices change daily, so check the source before quoting. Cheapest observed input price: $0.03/M.
| Provider | Model ref | Context | Quant | $/M in | $/M out | tok/s | TTFT | Tools | Status | Observed |
|---|---|---|---|---|---|---|---|---|---|---|
| DeepInfra deepinfra | Qwen/Qwen3-Reranker-4B | 32K | — | $0.03 | — | — | — | — | live | 2026-10-26 |
Articles and coverage
4 pieces
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.
Official1
docs.nvidia.com271 words
NVIDIA NeMo AutoModel loads Qwen3-Reranker-4B via Qwen3RerankerForCausalReranking to score query-document pairs using a pretrained language-model head. Saved checkpoints identify as standard Qwen3ForCausalLM models for loading in stock Transformers or vLLM. The reranker does not declare support for tensor, context, pipeline, or expert parallelism. The provided example launches on 8 GPUs.written, reviewed, cited
Analysis2
reranker.ukabout the Qwen3 Reranker family762 words
Qwen3-Reranker offers Apache 2.0 rerankers in 0.6B, 4B, and 8B sizes with a 32K context and 100+ languages. The 4B and 8B variants lead on MTEB-R and FollowIR suites across English, Chinese, multilingual, long-document, and code domains. Engineers should validate on their own labels before replacing bge or hosted APIs, noting that GPU resources are required for comfortable operation.written, reviewed, cited
- The 8B variant leads on Chinese, multilingual, long-document, and code tasks by at most 1.5 points.
- The raw transformers path requires Qwen's exact prompt template, otherwise quality drops without an error.
- The 4B model achieves the best FollowIR score in Qwen's table among the available sizes.
dev.co900 words
Qwen3-Reranker-4B is a 4-billion-parameter text reranking model from Alibaba designed to score and rank candidate documents against queries. The 36-layer model requires approximately 8–16 GB VRAM for fp32 inference and achieves ~10–20ms latency per pair on GPU hardware. With over 2M downloads and Apache 2.0 licensing, it supports 100+ languages and 32k token contexts for production retrieval augmentation.written, reviewed, cited
- The model outputs logit-based scores representing relevance rather than probabilities by default.
- Fine-tuning feasibility is unknown and likely requires a custom training pipeline with substantial labeled data.
- The model does not produce fixed-dimension vectors and cannot be used for semantic embedding generation.
Community1
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-Embedding-4B · 4.02B same size band (M), permissive license
- Qwen3-Embedding-8B · 7.57B same size band (M), more providers, permissive license
- CLM-v0.1-8B same task (text-ranking), permissive license
Sources
3 sources
Sources
Hub snapshot: 2026-10-26 · HF API: 2026-10-26 07:16 UTC · Providers: 2026-10-26 07:26 UTC · Articles: 2026-10-26 07:27 UTC · Summary: 2026-10-07 15:01 UTC · Last verified 2026-10-26 07:27 UTC
Follow Qwen3-Reranker-4B
Watch this model
- /feeds/models/Qwen--Qwen3-Reranker-4B.xml(RSS 2.0) · every change to Qwen3-Reranker-4B
- /feeds/families/Qwen--Qwen3-Reranker.xml(RSS 2.0) · every change across the Qwen3-Reranker family
Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.