Jamjuri Preview

Jamjuri β€” Preview

Jamjuri is the first fine-tuned model released by Cheva Labs.

Target Hub repository: Cheva123/Jamjuri-Preview-9B
Published artifact: JamjuriLLM-9B-Q8_0.gguf only

Series: part of Jamjuri 9B β€” collection Β· series card Series: part of the Jamjuri 9B series Β· collection

This preview is an early experiment focused on practical Thai SME workflows: Thai communication, document drafting, RAG, tool use, and simple business automation.

Experimental preview β€” do not use this model for production or real business operations.
Jamjuri is still in a very early stage of development. Its outputs may be incomplete, incorrect, inconsistent, or unsafe for real-world decisions. Do not rely on it for financial, legal, medical, operational, or other high-impact work.

English

What this preview is

Jamjuri Preview is Cheva Labs' first public fine-tuned model. The goal of this release is not to present a finished production model, but to document an early step toward a Thai-first local AI assistant for small and medium-sized businesses.

Current areas of focus:

  • Thai workplace communication
  • Document and business writing
  • Retrieval-Augmented Generation (RAG)
  • Tool / function calling
  • Simple workflow automation
  • Preserving general reasoning, math, coding, and instruction-following ability

Model

  • Base model: Qwen3.5-9B
  • Post-training: 3 Expert LoRA + merge
  • Primary preview format: GGUF Q8_0
  • Languages: Thai and English
  • Status: Experimental research preview

Early benchmark snapshot

The table below is an early, vendor-selected evaluation subset used during development. It is not a complete benchmark suite, does not measure 100% of model capability, and should not be treated as a universal ranking.

Benchmark Qwen Q8 Jamjuri Q8 Difference
BFCL V4 β€” simple_python 94.00% 93.00% -1.00 pp
IFEval β€” strict 83.36% 84.10% +0.74 pp
Thai Eval β€” Belebele fallback 75.00% 80.78% +5.78 pp
GSM8K β€” strict 83.40% 84.84% +1.44 pp
MMLU-Pro β€” 700-sample gate 72.57% 73.29% +0.71 pp
HumanEval+ β€” pass@1 82.90% 84.80% +1.90 pp
MBPP+ β€” pass@1 65.90% 64.30% -1.60 pp

In this development snapshot, Jamjuri scored higher on 5 of the 7 selected no-thinking evaluations. These results are useful for internal comparison, but they are not sufficient to claim broad superiority over the base model.

τ³ Retail result

A thinking-mode τ³ Retail run completed 114/114 terminal simulations for both models. Mean reward on reward-bearing records was 15.91% for Qwen (7/44) and 24.07% for Jamjuri (13/54). However, the two runs did not use an identical external judge route, so this result is treated as directional evidence only, not a controlled apples-to-apples benchmark. Infrastructure errors were recorded in both runs.

A broader set of standard and community benchmarks will be published later after the evaluation setup is made more reproducible and complete.

Intended use

This model is intended for:

  • Research and experimentation
  • Local inference testing
  • Thai SME workflow prototyping
  • RAG / tool-use experiments
  • Benchmarking and evaluation

Not intended for

Do not use this preview as a production business assistant or autonomous decision-maker.

Avoid using it for:

  • Financial decisions or accounting approval
  • Legal advice or legal document approval
  • Medical or safety-critical decisions
  • Autonomous purchases, payments, or irreversible actions
  • Production customer support without human review
  • Any workflow where an incorrect answer can cause material harm

Release philosophy

Jamjuri Preview is being released early so the work can be documented, tested, and improved in public. The model is expected to change substantially in later versions.

ΰΈ ΰΈ²ΰΈ©ΰΈ²ΰΉ„ΰΈ—ΰΈ’

Jamjuri Preview ΰΈ„ΰΈ·ΰΈ­ΰΈ­ΰΈ°ΰΉ„ΰΈ£

Jamjuri Preview ΰΉ€ΰΈ›ΰΉ‡ΰΈ™ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ Fine-tune ΰΈ•ΰΈ±ΰΈ§ΰΉΰΈ£ΰΈΰΈ—ΰΈ΅ΰΉˆ Cheva Labs ΰΈ›ΰΈ₯่อฒΰΈͺู่ΰΈͺΰΈ²ΰΈ˜ΰΈ²ΰΈ£ΰΈ“ΰΈ° ΰΉ‚ΰΈ”ΰΈ’ΰΉ€ΰΈ›ΰΉ‰ΰΈ²ΰΈ«ΰΈ‘ΰΈ²ΰΈ’ΰΈ‚ΰΈ­ΰΈ‡ΰΈ£ΰΈΈΰΉˆΰΈ™ΰΈ™ΰΈ΅ΰΉ‰ΰΉ„ΰΈ‘ΰΉˆΰΉƒΰΈŠΰΉˆΰΈΰΈ²ΰΈ£ΰΉ€ΰΈ›ΰΉ‡ΰΈ™ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ΰΈžΰΈ£ΰΉ‰ΰΈ­ΰΈ‘ΰΉƒΰΈŠΰΉ‰ΰΈ‡ΰΈ²ΰΈ™ΰΈˆΰΈ£ΰΈ΄ΰΈ‡ ΰΉΰΈ•ΰΉˆΰΉ€ΰΈ›ΰΉ‡ΰΈ™ΰΈΰΉ‰ΰΈ²ΰΈ§ΰΉΰΈ£ΰΈΰΈ‚ΰΈ­ΰΈ‡ΰΈΰΈ²ΰΈ£ΰΈ—ΰΈ”ΰΈ₯ΰΈ­ΰΈ‡ΰΈžΰΈ±ΰΈ’ΰΈ™ΰΈ² Local AI ΰΈ—ΰΈ΅ΰΉˆΰΉ€ΰΈ™ΰΉ‰ΰΈ™ΰΈΰΈ²ΰΈ£ΰΉƒΰΈŠΰΉ‰ΰΈ‡ΰΈ²ΰΈ™ΰΉƒΰΈ™ΰΈšΰΈ£ΰΈ΄ΰΈšΰΈ—ΰΈ‚ΰΈ­ΰΈ‡ SME ΰΉ„ΰΈ—ΰΈ’

ΰΈ„ΰΈ§ΰΈ²ΰΈ‘ΰΈͺΰΈ²ΰΈ‘ΰΈ²ΰΈ£ΰΈ–ΰΈ—ΰΈ΅ΰΉˆΰΈΰΈ³ΰΈ₯ΰΈ±ΰΈ‡ΰΈžΰΈ±ΰΈ’ΰΈ™ΰΈ²ΰΉƒΰΈ™ΰΈ£ΰΈΈΰΉˆΰΈ™ΰΈ™ΰΈ΅ΰΉ‰ ΰΉ„ΰΈ”ΰΉ‰ΰΉΰΈΰΉˆ

  • การΰΈͺื่อΰΈͺΰΈ²ΰΈ£ΰΈ ΰΈ²ΰΈ©ΰΈ²ΰΉ„ΰΈ—ΰΈ’ΰΉƒΰΈ™ΰΈšΰΈ£ΰΈ΄ΰΈšΰΈ—ΰΈΰΈ²ΰΈ£ΰΈ—ΰΈ³ΰΈ‡ΰΈ²ΰΈ™
  • การเขมฒนเอกΰΈͺารแΰΈ₯ΰΈ°ΰΈ‚ΰΉ‰ΰΈ­ΰΈ„ΰΈ§ΰΈ²ΰΈ‘ΰΈ—ΰΈ²ΰΈ‡ΰΈ˜ΰΈΈΰΈ£ΰΈΰΈ΄ΰΈˆ
  • Retrieval-Augmented Generation (RAG)
  • Tool / Function Calling
  • Workflow automation ΰΉΰΈšΰΈšΰΈžΰΈ·ΰΉ‰ΰΈ™ΰΈΰΈ²ΰΈ™
  • การรักษาควาฑΰΈͺΰΈ²ΰΈ‘ΰΈ²ΰΈ£ΰΈ–ΰΉ€ΰΈ”ΰΈ΄ΰΈ‘ΰΈ”ΰΉ‰ΰΈ²ΰΈ™ Reasoning, Math, Coding แΰΈ₯ΰΈ° Instruction Following

ΰΈ‚ΰΉ‰ΰΈ­ΰΈ‘ΰΈΉΰΈ₯ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯

  • Base model: Qwen3.5-9B
  • Post-training: 3 Expert LoRA + Merge
  • ΰΈ£ΰΈΉΰΈ›ΰΉΰΈšΰΈš Preview ΰΈ«ΰΈ₯ัก: GGUF Q8_0
  • ΰΈ ΰΈ²ΰΈ©ΰΈ²: ไทฒแΰΈ₯ะอังก฀ษ
  • ΰΈͺΰΈ–ΰΈ²ΰΈ™ΰΈ°: Experimental Research Preview

ΰΉ„ΰΈ‘ΰΉˆΰΈ„ΰΈ§ΰΈ£ΰΈ™ΰΈ³ Jamjuri Preview ΰΉ„ΰΈ›ΰΉƒΰΈŠΰΉ‰ΰΉƒΰΈ™ΰΈ‡ΰΈ²ΰΈ™ΰΈˆΰΈ£ΰΈ΄ΰΈ‡ΰΈ«ΰΈ£ΰΈ·ΰΈ­ΰΈ£ΰΈ°ΰΈšΰΈš Production
ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ΰΈ’ΰΈ±ΰΈ‡ΰΈ­ΰΈ’ΰΈΉΰΉˆΰΉƒΰΈ™ΰΈŠΰΉˆΰΈ§ΰΈ‡ΰΈ—ΰΈ”ΰΈ₯องระฒะแรกฑาก ΰΈ„ΰΈ³ΰΈ•ΰΈ­ΰΈšΰΈ­ΰΈ²ΰΈˆΰΈœΰΈ΄ΰΈ” ΰΉ„ΰΈ‘ΰΉˆΰΈ„ΰΈ£ΰΈš ΰΉ„ΰΈ‘ΰΉˆΰΈͺΰΈ‘ΰΉˆΰΈ³ΰΉ€ΰΈͺΰΈ‘ΰΈ­ ΰΈ«ΰΈ£ΰΈ·ΰΈ­ΰΈ‘ΰΈ΅ΰΈžΰΈ€ΰΈ•ΰΈ΄ΰΈΰΈ£ΰΈ£ΰΈ‘ΰΈ—ΰΈ΅ΰΉˆΰΈ’ΰΈ±ΰΈ‡ΰΉ„ΰΈ‘ΰΉˆΰΉ„ΰΈ”ΰΉ‰ΰΈ£ΰΈ±ΰΈšΰΈΰΈ²ΰΈ£ΰΈ—ΰΈ”ΰΈͺΰΈ­ΰΈšΰΉ€ΰΈžΰΈ΅ΰΈ’ΰΈ‡ΰΈžΰΈ­ ΰΈ—ΰΈΈΰΈΰΈ‡ΰΈ²ΰΈ™ΰΈ—ΰΈ΅ΰΉˆΰΈ‘ΰΈ΅ΰΈœΰΈ₯ΰΈΰΈ£ΰΈ°ΰΈ—ΰΈšΰΈˆΰΈ£ΰΈ΄ΰΈ‡ΰΈ„ΰΈ§ΰΈ£ΰΈ‘ΰΈ΅ΰΈ‘ΰΈ™ΰΈΈΰΈ©ΰΈ’ΰΉŒΰΈ•ΰΈ£ΰΈ§ΰΈˆΰΈͺΰΈ­ΰΈšΰΉ€ΰΈͺΰΈ‘ΰΈ­

ผΰΈ₯ Benchmark ΰΉ€ΰΈšΰΈ·ΰΉ‰ΰΈ­ΰΈ‡ΰΈ•ΰΉ‰ΰΈ™

ผΰΈ₯ΰΈ”ΰΉ‰ΰΈ²ΰΈ™ΰΈ₯ΰΉˆΰΈ²ΰΈ‡ΰΉ€ΰΈ›ΰΉ‡ΰΈ™ ΰΈŠΰΈΈΰΈ” Benchmark ΰΈ—ΰΈ΅ΰΉˆ Cheva Labs ΰΉ€ΰΈ₯ΰΈ·ΰΈ­ΰΈΰΉƒΰΈŠΰΉ‰ΰΈ£ΰΈ°ΰΈ«ΰΈ§ΰΉˆΰΈ²ΰΈ‡ΰΈΰΈ²ΰΈ£ΰΈžΰΈ±ΰΈ’ΰΈ™ΰΈ² ΰΉ€ΰΈ—ΰΉˆΰΈ²ΰΈ™ΰΈ±ΰΉ‰ΰΈ™ ΰΉ„ΰΈ‘ΰΉˆΰΉƒΰΈŠΰΉˆΰΈŠΰΈΈΰΈ”ΰΈ—ΰΈ”ΰΈͺΰΈ­ΰΈšΰΈ—ΰΈ΅ΰΉˆΰΈ„ΰΈ£ΰΈ­ΰΈšΰΈ„ΰΈ₯ΰΈΈΰΈ‘ΰΈ„ΰΈ§ΰΈ²ΰΈ‘ΰΈͺΰΈ²ΰΈ‘ΰΈ²ΰΈ£ΰΈ–ΰΈ—ΰΈ±ΰΉ‰ΰΈ‡ΰΈ«ΰΈ‘ΰΈ”ΰΈ‚ΰΈ­ΰΈ‡ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ แΰΈ₯ΰΈ° ΰΉ„ΰΈ‘ΰΉˆΰΉ„ΰΈ”ΰΉ‰ΰΈ«ΰΈ‘ΰΈ²ΰΈ’ΰΈ„ΰΈ§ΰΈ²ΰΈ‘ΰΈ§ΰΉˆΰΈ²ΰΉ€ΰΈ›ΰΉ‡ΰΈ™ΰΈœΰΈ₯ Benchmark ΰΈͺากΰΈ₯แบบ 100%

Benchmark Qwen Q8 Jamjuri Q8 ΰΈ•ΰΉˆΰΈ²ΰΈ‡ΰΈΰΈ±ΰΈ™
BFCL V4 β€” simple_python 94.00% 93.00% -1.00 pp
IFEval β€” strict 83.36% 84.10% +0.74 pp
Thai Eval β€” Belebele fallback 75.00% 80.78% +5.78 pp
GSM8K β€” strict 83.40% 84.84% +1.44 pp
MMLU-Pro β€” 700-sample gate 72.57% 73.29% +0.71 pp
HumanEval+ β€” pass@1 82.90% 84.80% +1.90 pp
MBPP+ β€” pass@1 65.90% 64.30% -1.60 pp

ΰΉƒΰΈ™ΰΈŠΰΈΈΰΈ”ΰΈ—ΰΈ”ΰΈͺΰΈ­ΰΈšΰΉ€ΰΈšΰΈ·ΰΉ‰ΰΈ­ΰΈ‡ΰΈ•ΰΉ‰ΰΈ™ΰΉΰΈšΰΈš no-thinking ΰΈ™ΰΈ΅ΰΉ‰ Jamjuri ฑมคะแนนΰΈͺΰΈΉΰΈ‡ΰΈΰΈ§ΰΉˆΰΈ² Base ΰΉƒΰΈ™ 5 จาก 7 ราฒการ ΰΉΰΈ•ΰΉˆΰΈœΰΈ₯นม้ควรถูกฑองเป็นข้อฑูΰΈ₯ΰΈͺΰΈ³ΰΈ«ΰΈ£ΰΈ±ΰΈšΰΈΰΈ²ΰΈ£ΰΉ€ΰΈ›ΰΈ£ΰΈ΅ΰΈ’ΰΈšΰΉ€ΰΈ—ΰΈ΅ΰΈ’ΰΈšΰΈ£ΰΈ°ΰΈ«ΰΈ§ΰΉˆΰΈ²ΰΈ‡ΰΈΰΈ²ΰΈ£ΰΈžΰΈ±ΰΈ’ΰΈ™ΰΈ²ΰΉ€ΰΈ—ΰΉˆΰΈ²ΰΈ™ΰΈ±ΰΉ‰ΰΈ™ แΰΈ₯ΰΈ°ΰΈ’ΰΈ±ΰΈ‡ΰΉ„ΰΈ‘ΰΉˆΰΉ€ΰΈžΰΈ΅ΰΈ’ΰΈ‡ΰΈžΰΈ­ΰΈͺำหรับΰΈͺΰΈ£ΰΈΈΰΈ›ΰΈ§ΰΉˆΰΈ²ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ΰΈ”ΰΈ΅ΰΈΰΈ§ΰΉˆΰΈ² Base ΰΉƒΰΈ™ΰΈ ΰΈ²ΰΈžΰΈ£ΰΈ§ΰΈ‘ΰΈ—ΰΈΈΰΈΰΈ”ΰΉ‰ΰΈ²ΰΈ™

τ³ Retail

การทดΰΈͺอบ τ³ Retail แบบ thinking-mode ΰΉ€ΰΈͺΰΈ£ΰΉ‡ΰΈˆΰΈͺΰΈ΄ΰΉ‰ΰΈ™ 114/114 terminal simulations ΰΈͺΰΈ³ΰΈ«ΰΈ£ΰΈ±ΰΈšΰΈ—ΰΈ±ΰΉ‰ΰΈ‡ΰΈͺΰΈ­ΰΈ‡ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ ΰΉ‚ΰΈ”ΰΈ’ΰΈ‘ΰΈ΅ΰΈ„ΰΉˆΰΈ² mean reward ΰΈˆΰΈ²ΰΈΰΈ£ΰΈ²ΰΈ’ΰΈΰΈ²ΰΈ£ΰΈ—ΰΈ΅ΰΉˆΰΈ‘ΰΈ΅ reward ΰΉ€ΰΈ—ΰΉˆΰΈ²ΰΈΰΈ±ΰΈš 15.91% ΰΈͺำหรับ Qwen (7/44) แΰΈ₯ΰΈ° 24.07% ΰΈͺำหรับ Jamjuri (13/54) ΰΈ­ΰΈ’ΰΉˆΰΈ²ΰΈ‡ΰΉ„ΰΈ£ΰΈΰΉ‡ΰΈ•ΰΈ²ΰΈ‘ ΰΈ£ΰΈ­ΰΈšΰΈ—ΰΈ”ΰΈͺΰΈ­ΰΈšΰΈ‚ΰΈ­ΰΈ‡ΰΈͺΰΈ­ΰΈ‡ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ΰΉ„ΰΈ‘ΰΉˆΰΉ„ΰΈ”ΰΉ‰ΰΉƒΰΈŠΰΉ‰ external judge route ΰΈ—ΰΈ΅ΰΉˆΰΉ€ΰΈ«ΰΈ‘ΰΈ·ΰΈ­ΰΈ™ΰΈΰΈ±ΰΈ™ΰΈ—ΰΈΈΰΈΰΈ›ΰΈ£ΰΈ°ΰΈΰΈ²ΰΈ£ ΰΈ”ΰΈ±ΰΈ‡ΰΈ™ΰΈ±ΰΉ‰ΰΈ™ΰΈœΰΈ₯ΰΈ™ΰΈ΅ΰΉ‰ΰΈ–ΰΈ·ΰΈ­ΰΉ€ΰΈ›ΰΉ‡ΰΈ™ directional evidence ΰΉ€ΰΈ—ΰΉˆΰΈ²ΰΈ™ΰΈ±ΰΉ‰ΰΈ™ ΰΉ„ΰΈ‘ΰΉˆΰΉƒΰΈŠΰΉˆ controlled benchmark ΰΈ—ΰΈ΅ΰΉˆΰΈͺΰΈ‘ΰΈšΰΈΉΰΈ£ΰΈ“ΰΉŒ แΰΈ₯ΰΈ°ΰΈ—ΰΈ±ΰΉ‰ΰΈ‡ΰΈͺΰΈ­ΰΈ‡ΰΈ£ΰΈ­ΰΈšΰΈ‘ΰΈ΅ infrastructure error ΰΈšΰΈ²ΰΈ‡ΰΈͺΰΉˆΰΈ§ΰΈ™

Benchmark ΰΈͺากΰΈ₯แΰΈ₯ΰΈ° Community Benchmark ΰΈ—ΰΈ΅ΰΉˆΰΈΰΈ§ΰΉ‰ΰΈ²ΰΈ‡ΰΈΰΈ§ΰΉˆΰΈ²ΰΈ™ΰΈ΅ΰΉ‰ΰΈˆΰΈ°ΰΈ—ΰΈ’ΰΈ­ΰΈ’ΰΉ€ΰΈžΰΈ΄ΰΉˆΰΈ‘ΰΉƒΰΈ™ΰΈ ΰΈ²ΰΈ’ΰΈ«ΰΈ₯ΰΈ±ΰΈ‡ ΰΈ«ΰΈ₯ΰΈ±ΰΈ‡ΰΈˆΰΈ²ΰΈΰΈ£ΰΈ°ΰΈšΰΈšΰΈ›ΰΈ£ΰΈ°ΰΉ€ΰΈ‘ΰΈ΄ΰΈ™ΰΈ‘ΰΈ΅ΰΈ„ΰΈ§ΰΈ²ΰΈ‘ reproducible แΰΈ₯ΰΈ°ΰΈ„ΰΈ£ΰΈ­ΰΈšΰΈ„ΰΈ₯ุฑฑากขย้น

ΰΉ€ΰΈ«ΰΈ‘ΰΈ²ΰΈ°ΰΈͺำหรับ

  • ΰΈ‡ΰΈ²ΰΈ™ΰΈ§ΰΈ΄ΰΈˆΰΈ±ΰΈ’ΰΉΰΈ₯ะการทดΰΈ₯ΰΈ­ΰΈ‡
  • การทดΰΈͺอบ Local inference
  • Prototype workflow ΰΈͺำหรับ SME ΰΉ„ΰΈ—ΰΈ’
  • การทดΰΈ₯ΰΈ­ΰΈ‡ RAG แΰΈ₯ΰΈ° Tool Use
  • Benchmark แΰΈ₯ΰΈ° Evaluation

ΰΉ„ΰΈ‘ΰΉˆΰΉ€ΰΈ«ΰΈ‘ΰΈ²ΰΈ°ΰΈͺำหรับ

ΰΉ„ΰΈ‘ΰΉˆΰΈ„ΰΈ§ΰΈ£ΰΉƒΰΈŠΰΉ‰ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ΰΈ™ΰΈ΅ΰΉ‰ΰΉ€ΰΈ›ΰΉ‡ΰΈ™ΰΈœΰΈΉΰΉ‰ΰΈŠΰΉˆΰΈ§ΰΈ’ΰΈ˜ΰΈΈΰΈ£ΰΈΰΈ΄ΰΈˆΰΉƒΰΈ™ Production ΰΈ«ΰΈ£ΰΈ·ΰΈ­ΰΈ£ΰΈ°ΰΈšΰΈšΰΈ•ΰΈ±ΰΈ”ΰΈͺΰΈ΄ΰΈ™ΰΉƒΰΈˆΰΈ­ΰΈ±ΰΈ•ΰΉ‚ΰΈ™ΰΈ‘ΰΈ±ΰΈ•ΰΈ΄

ΰΈ„ΰΈ§ΰΈ£ΰΈ«ΰΈ₯มกเΰΈ₯ΰΈ΅ΰΉˆΰΈ’ΰΈ‡ΰΈΰΈ²ΰΈ£ΰΉƒΰΈŠΰΉ‰ΰΈΰΈ±ΰΈš:

  • การตัดΰΈͺΰΈ΄ΰΈ™ΰΉƒΰΈˆΰΈ”ΰΉ‰ΰΈ²ΰΈ™ΰΈΰΈ²ΰΈ£ΰΉ€ΰΈ‡ΰΈ΄ΰΈ™ΰΈ«ΰΈ£ΰΈ·ΰΈ­ΰΈ­ΰΈ™ΰΈΈΰΈ‘ΰΈ±ΰΈ•ΰΈ΄ΰΈšΰΈ±ΰΈΰΈŠΰΈ΅
  • คำแนะนำหรือการอนุฑัติเอกΰΈͺΰΈ²ΰΈ£ΰΈ”ΰΉ‰ΰΈ²ΰΈ™ΰΈΰΈŽΰΈ«ΰΈ‘ΰΈ²ΰΈ’
  • ΰΈΰΈ²ΰΈ£ΰΉΰΈžΰΈ—ΰΈ’ΰΉŒΰΈ«ΰΈ£ΰΈ·ΰΈ­ΰΈ£ΰΈ°ΰΈšΰΈšΰΈ—ΰΈ΅ΰΉˆΰΉ€ΰΈΰΈ΅ΰΉˆΰΈ’ΰΈ§ΰΈ‚ΰΉ‰ΰΈ­ΰΈ‡ΰΈΰΈ±ΰΈšΰΈ„ΰΈ§ΰΈ²ΰΈ‘ΰΈ›ΰΈ₯ΰΈ­ΰΈ”ΰΈ ΰΈ±ΰΈ’
  • การซื้อ ΰΈˆΰΉˆΰΈ²ΰΈ’ΰΉ€ΰΈ‡ΰΈ΄ΰΈ™ ΰΈ«ΰΈ£ΰΈ·ΰΈ­ΰΈ—ΰΈ³ΰΈ„ΰΈ³ΰΈͺΰΈ±ΰΉˆΰΈ‡ΰΈ—ΰΈ΅ΰΉˆΰΈ’ΰΉ‰ΰΈ­ΰΈ™ΰΈΰΈ₯ΰΈ±ΰΈšΰΉ„ΰΈ‘ΰΉˆΰΉ„ΰΈ”ΰΉ‰ΰΉΰΈšΰΈšΰΈ­ΰΈ±ΰΈ•ΰΉ‚ΰΈ™ΰΈ‘ΰΈ±ΰΈ•ΰΈ΄
  • Customer support ΰΈˆΰΈ£ΰΈ΄ΰΈ‡ΰΉ‚ΰΈ”ΰΈ’ΰΉ„ΰΈ‘ΰΉˆΰΈ‘ΰΈ΅ΰΈ‘ΰΈ™ΰΈΈΰΈ©ΰΈ’ΰΉŒΰΈ•ΰΈ£ΰΈ§ΰΈˆΰΈͺอบ
  • ΰΈ‡ΰΈ²ΰΈ™ΰΉƒΰΈ” ΰΉ† ΰΈ—ΰΈ΅ΰΉˆΰΈ„ΰΈ³ΰΈ•ΰΈ­ΰΈšΰΈœΰΈ΄ΰΈ”ΰΈͺΰΈ²ΰΈ‘ΰΈ²ΰΈ£ΰΈ–ΰΈͺΰΈ£ΰΉ‰ΰΈ²ΰΈ‡ΰΈ„ΰΈ§ΰΈ²ΰΈ‘ΰΉ€ΰΈͺΰΈ΅ΰΈ’ΰΈ«ΰΈ²ΰΈ’ΰΈˆΰΈ£ΰΈ΄ΰΈ‡

แนวทางของ Preview

Jamjuri Preview ถูกปΰΈ₯ΰΉˆΰΈ­ΰΈ’ΰΉƒΰΈ™ΰΈŠΰΉˆΰΈ§ΰΈ‡ΰΉΰΈ£ΰΈΰΉ€ΰΈžΰΈ·ΰΉˆΰΈ­ΰΉƒΰΈ«ΰΉ‰ΰΈͺΰΈ²ΰΈ‘ΰΈ²ΰΈ£ΰΈ–ΰΈšΰΈ±ΰΈ™ΰΈ—ΰΈΆΰΈΰΈœΰΈ₯ ΰΈ—ΰΈ”ΰΈ₯ΰΈ­ΰΈ‡ แΰΈ₯ΰΈ°ΰΈžΰΈ±ΰΈ’ΰΈ™ΰΈ²ΰΈ•ΰΉˆΰΈ­ΰΉΰΈšΰΈšΰΉ€ΰΈ›ΰΈ΄ΰΈ”ΰΉ€ΰΈœΰΈ’ΰΉ„ΰΈ”ΰΉ‰ ΰΉ‚ΰΈ‘ΰΉ€ΰΈ”ΰΈ₯ΰΈ£ΰΈΈΰΉˆΰΈ™ΰΈ–ΰΈ±ΰΈ”ΰΉ„ΰΈ›ΰΈ­ΰΈ²ΰΈˆΰΉ€ΰΈ›ΰΈ₯ΰΈ΅ΰΉˆΰΈ’ΰΈ™ΰΉΰΈ›ΰΈ₯ΰΈ‡ΰΈˆΰΈ²ΰΈΰΈ£ΰΈΈΰΉˆΰΈ™ΰΈ™ΰΈ΅ΰΉ‰ΰΈ­ΰΈ’ΰΉˆΰΈ²ΰΈ‡ΰΈ‘ΰΈ²ΰΈ

Local inference

llama.cpp

llama-server \
  --model JamjuriLLM-9B-Q8_0.gguf \
  --jinja \
  --ctx-size 16384 \
  --n-predict 4096 \
  --temperature 0.7 \
  --host 127.0.0.1 \
  --port 8080

The OpenAI-compatible endpoint is available at http://127.0.0.1:8080/v1.

Ollama

ollama create jamjuri:9b-q8 -f Modelfile.Q8_0
ollama run jamjuri:9b-q8

Release files

  • JamjuriLLM-9B-Q8_0.gguf β€” the only model weight staged for this release
  • Modelfile.Q8_0 β€” Ollama configuration
  • assets/jamjuri-preview.png β€” model-card artwork
  • MANIFEST.json β€” local checksum and provenance record

This preview uses the Apache-2.0 license, matching the Qwen3.5-9B base model. Dataset and adapter redistribution terms should still be reviewed separately before public release.

Cheva Labs

Cheva Labs builds open, practical AI and engineering tools with an emphasis on local AI, reproducible experiments, and making advanced systems more accessible to Thai users, students, individuals, and smaller organizations.

Downloads last month
12
GGUF
Model size
9B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for Cheva123/Jamjuri-Preview-9B

Finetuned
Qwen/Qwen3.5-9B
Quantized
(743)
this model

Collection including Cheva123/Jamjuri-Preview-9B