Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Built upon extensive training, Qwen3 delivers groundbreaking advancements in reasoning, instruction-following, agent capabilities, and multilingual support, with the following key features: - Uniquely support of seamless switching between thinking mode (for complex logical reasoning, math, and coding) and non-thinking mode (for efficient, general-purpose dialogue) within single model, ensuring optimal performance across various scenarios. - Significantly enhancement in its reasoning capabilities, surpassing previous QwQ (in thinking mode) and…
kas-4b is an open-weight model for text generation from Kashav Piya, released under Apache License 2.0. It has 4B parameters and a 40,960-token context. At 16-bit it needs about 9.7 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 46 downloads a month.
A fine-tuned decision engine based on Qwen/Qwen3-4B, submitted to the Decision Index v0.2.1 leaderboard. Decision Index: 40.1 (raw 54.18) — above Tev1-4B (29.2), within 6 points of Scion v4 9B (45.9).
Parameters4B
Context40,960
Weights8.0 GB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads46
Runs On
What it takes to serve kas-4b (4B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 8.0 GB | 9.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 4.0 GB | 4.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 2.0 GB | 2.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 7, 2026.
kas-4b on every accelerator the SAVRN Index prices, at every precision
Model Card
By Kashav Piya, published under apache-2.0, revision eb59930976ee.
A fine-tuned decision engine based on Qwen/Qwen3-4B, submitted to the Decision Index v0.2.1 leaderboard. Decision Index: 40.1 (raw 54.18) — above Tev1-4B (29.2), within 6 points of Scion v4 9B (45.9). Per-area breakdown against all 4B submissions on Decision Index 0.2.1. kas-4b leads the 4B tier on: - Knowledge & Reasoning (39.5) — highest of all 4B models; higher LoRA rank (r64) likely helps on harder reasoning tasks - Arts & Human Taste (37.7) — highest of all 4B models by a wide margin - Language Understanding (35.2) — 13–25 points behind ezjev and Nox; language-heavy fine-tuning data favors those models - Tools & Automation (49.0) — behind ezjev (69.9) and Nox (60.1) Profile: the most…
Read Kashav Piya's full model card
A fine-tuned decision engine based on Qwen/Qwen3-4B, submitted to the Decision Index v0.2.1 leaderboard.
Decision Index: 40.1 (raw 54.18) — above Tev1-4B (29.2), within 6 points of Scion v4 9B (45.9).
Scores
| Area | Score |
|---|---|
| Tools & Automation | 49.0 |
| Retrieval & Classification | 40.4 |
| Knowledge & Reasoning | 39.5 |
| Arts & Human Taste | 37.7 |
| Language Understanding | 35.2 |
| Decision Index | 40.1 |
Full results: kpiya/decision-index-results
4B Model Comparison — Where kas-4b Leads and Trails
Per-area breakdown against all 4B submissions on Decision Index 0.2.1.
| Area | ezjev-4b (51.2) | Nox 4B (43.8) | kas-4b (40.1) | intelif-4B (31.8) |
|---|---|---|---|---|
| Tools & Automation | 69.9 | 60.1 | 49.0 | 51.0 |
| Retrieval & Classification | 56.3 | 52.4 | 40.4 | 39.6 |
| Knowledge & Reasoning | 33.5 | 27.6 | 39.5 | 18.3 |
| Language Understanding | 60.2 | 48.6 | 35.2 | 31.1 |
| Arts & Human Taste | 28.8 | 25.8 | 37.7 | 17.8 |
kas-4b leads the 4B tier on: - Knowledge & Reasoning (39.5) — highest of all 4B models; higher LoRA rank (r64) likely helps on harder reasoning tasks - Arts & Human Taste (37.7) — highest of all 4B models by a wide margin
kas-4b trails on: - Language Understanding (35.2) — 13–25 points behind ezjev and Nox; language-heavy fine-tuning data favors those models - Tools & Automation (49.0) — behind ezjev (69.9) and Nox (60.1)
Profile: the most balanced 4B model on the benchmark — consistent across all five areas rather than peaking on one or two.
Training
- Base model: Qwen/Qwen3-4B (Apache 2.0, pure transformer)
- Method: LoRA fine-tune (PEFT), adapter merged into base weights
- LoRA config: rank 64, alpha 128, 2 epochs, LR 1e-4, batch size 16, 4096 token limit
- Recipe: inspired by Scion (Sinan Ozdemir)
- Hardware: NVIDIA RTX PRO 6000 Blackwell (HF Jobs)
Engine
This model is served via a custom engine (kas_engine.engine:KasEngine) that:
- Maps options to single-token uppercase labels (A–Z, then two-letter)
- Repeats the JSON prompt once (Scion-style)
- Scores label logits at the generation position, softmaxed to probabilities
- Refuses over-context prompts — never truncates
Engine code: kashavpiya/kas-4b
Latency
Measured on RTX PRO 6000, single process, one request at a time:
| Metric | Value |
|---|---|
| Median | 39.2 ms |
| p95 | 546.4 ms |
| Mean | 153.3 ms |
Evaluation
Run with Decision Index kit v0.2.1 on 150,759 rows (184 unsupported / over-context, not truncated).
Notes
Not trained on Decision Index suite data. Base model license: Apache 2.0.
Configuration
- Architecture
- Qwen3ForCausalLM
- Context length (tokens)
- 40,960
- Layers
- 36
- Hidden size
- 2,560
- Feed-forward size
- 9,728
- Attention heads
- 32
- Key/value heads
- 8
- Head dimension
- 128
- Vocabulary size
- 151,936
- Model type
- qwen3
Identity and Version
- Repository
- kpiya/kas-4b
- Publisher
- Kashav Piya
- Task
- Text generation
- Modality
- Text
- Library
- Not stated by the source
- Parameters
- 4B parameters
- Languages
- en
- Revision
- eb59930976ee392cb0557528a7520935f33a386e
- First published
- 2026-10-04
- Last updated
- 2026-10-07
Files and Weights
8 files, 8.1 GB in total. The weights are 1 file totalling 8.0 GB in safetensors.
Weights1 file · 8.0 GB
Configuration2 files · 1.8 KB
Tokenizer2 files · 11.4 MB
Documentation1 file · 3.3 KB
Other1 file · 4.2 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 8.0 GB | 92cb90ec3db4 |
| config.json | Configuration | 1.6 KB | — |
| generation_config.json | Configuration | 214 B | — |
| README.md | Documentation | 3.3 KB | — |
| chat_template.jinja | Other | 4.2 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 11.4 MB | be75606093db |
| tokenizer_config.json | Tokenizer | 694 B | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 8.0 GB
Download from Kashav Piya
Released by Kashav Piya through its official repository on Hugging Face. Read the license.
Built From
- Adapter of Qwen/Qwen3-4B
- Derived from Qwen/Qwen3-4B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 8.0 GB |
| 16-bit | 8.0 GB |
| 8-bit | 4.0 GB |
| 4-bit | 2.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About kas-4b
How much GPU memory does kas-4b need?
About 9.7 GB at 16-bit and 2.4 GB at 4-bit: the weights (4B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run kas-4b on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use kas-4b commercially?
Yes. kas-4b is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.
What is kas-4b's context length?
40,960 tokens, from the maximum position embeddings in its published configuration.
Similar Models
We introduce the updated version of the Qwen3-4B non-thinking mode, named Qwen3-4B-Instruct-2507, featuring the following key enhancements: - Significant improvements in general capabilities, including instruction following, logical reasoning, text comprehension, mathematics, science, coding and tool usage. - Substantial gains in long-tail knowledge coverage across multiple languages. - Markedly better alignment with user preferences in subjective and open-ended tasks, enabling more helpful responses and higher-quality text generation. - Enhanced capabilities in 256K long-context understanding. Qwen3-4B-Instruct-2507 has the following features: NOTE: This model supports only non-thinking…
Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Building upon extensive advancements in training data, model architecture, and optimization techniques, Qwen3 delivers the following key improvements over the previously released Qwen2.5: Qwen3-4B-Base has the following features: For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation. The code of Qwen3 has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers. With transformers<4.51.0, you will…
Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models. Building upon extensive advancements in training data, model architecture, and optimization techniques, Qwen3 delivers the following key improvements over the previously released Qwen2.5: Qwen3-4B-Base has the following features: For more details, including benchmark evaluation, hardware requirements, and inference performance, please refer to our blog, GitHub, and Documentation. The code of Qwen3 has been in the latest Hugging Face transformers and we advise you to use the latest version of transformers. With transformers<4.51.0, you will…
SRA-RiskGate-4B is an autonomous risk scoring, compliance verification, and dispute adjudication model fine-tuned on top of Qwen/Qwen3-4B-Instruct-2507. It is engineered for both ends of a stablecoin payment's operational lifecycle: Autonomous on-chain agents can hallucinate payment transfers, sign malformed calldata, or trigger catastrophic transactions during market depegs. The deterministic pre-filter and policy firewall is available directly as a verified Action Provider for the Coinbase AgentKit framework: Register RiskGateActionProvider as the payment provider on your AgentKit instance. It checks chain support, policies, and peg deviations, executing safe ERC-20 transfers only after…
![[HYU_NLP] EVA Team logo](/media/mh-logo-hyu-nlp-eval-558e075b88.webp){kind=logo}
Intermediate policy from dynamic OnlineRubrics-Every GRPO training. Distinct from static-rubric GRPO. Base model: Qwen/Qwen3-4B-Instruct-2507; thinking disabled. This checkpoint is a policy state used by the Phase-1 audit. No downstream medical capability or safety claim is made. Research use only; not validated for clinical decision-making. Root files are the veRL-exported Hugging Face inference model (BF16). originalcheckpoint/ preserves the exact original FSDP parameter checkpoint and tokenizer/configuration files. Optimizer state, training data, responses, rubrics, infrastructure configuration, and credentials are not included. The original is retained because export…