Qwen 3.8 27B
Apache 2.0 Open SourceDeveloped by Alibaba Cloud · Dense Transformer · Modalities: Text, Code · Released 2026-02-15
What are the token costs and operational benchmarks for Qwen 3.8 27B?
Verified Capability Benchmarks
Composite Intelligence: 71.6/100Architectural & Production Profile
Compact 27B dense Apache 2.0 model engineered for high-accuracy code generation and single-GPU enterprise hosting.
High-performance 27B open weights optimized for local deployment, code generation, and low-latency production pipelines.
- Architecture
- Dense Transformer
- Modalities
- Text, Code
- Max Output Tokens
- 16,384 tokens
- Blended Rate (3:1)
- $0.300 / 1M
Verified Cloud Inference Endpoints Hosting Qwen 3.8 27B
Compare latency, streaming speed, and token pricing across cloud API providers.
| Inference Provider | Role | Input / 1M | Output / 1M | Avg TTFT | Throughput |
|---|---|---|---|---|---|
| Alibaba Cloud DashScope → | Primary Host | $0.20 | $0.60 | 120ms | 140 tok/s |
| Cerebras CS-3 Wafer-Scale Inference → | Partner Endpoint | $0.20 | $0.60 | 70ms | 392 tok/s |
Interactive Monthly Token Economics & ROI Forecaster
Model production workloads across prompt, completion, and cached context tokens, or switch models on the fly.
Estimated Monthly Spend
$0.20/1M in · $0.60/1M out
Prompt Cache: $0.050/1M read · $0.20/1M write
$2.00/1M in · $10.00/1M out
Prompt Cache: $0.100/1M read · $0.50/1M write
Head-to-Head Comparisons Involving Qwen 3.8 27B (36)
Qwen 3.8 27B vs Gemini 4 Argon
Qwen 3.8 27B is 90.0% cheaper for input tokens ($0.20 vs. $2.00 per 1M tokens) and $0.60 vs. $10.00 for output tokens (10.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (730ms faster than Gemini 4 Argon) with 140 vs. 115 tok/s throughput. Gemini 4 Argon leads coding evaluations at 84.5% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 93.8% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Claude 4.5 Sonnet
Qwen 3.8 27B is 93.3% cheaper for input tokens ($0.20 vs. $3.00 per 1M tokens) and $0.60 vs. $15.00 for output tokens (15.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (260ms faster than Claude 4.5 Sonnet) with 140 vs. 84 tok/s throughput. Claude 4.5 Sonnet leads coding evaluations at 75.8% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 84.2% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Claude Opus 5.5
Qwen 3.8 27B is 95.0% cheaper for input tokens ($0.20 vs. $4.00 per 1M tokens) and $0.60 vs. $20.00 for output tokens (20.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (560ms faster than Claude Opus 5.5) with 140 vs. 95 tok/s throughput. Claude Opus 5.5 leads coding evaluations at 89.9% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 93.2% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Gemini 3.8 Flash
Qwen 3.8 27B is 20.0% cheaper for input tokens ($0.20 vs. $0.25 per 1M tokens) and $0.60 vs. $1.50 for output tokens (1.2x cost advantage). In terms of operational performance, Gemini 3.8 Flash delivers lower response latency at 105ms TTFT (15ms faster than Qwen 3.8 27B) with 210 vs. 140 tok/s throughput. Gemini 3.8 Flash leads coding evaluations at 68.5% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 79.2% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs GPT-5
Qwen 3.8 27B is 92.0% cheaper for input tokens ($0.20 vs. $2.50 per 1M tokens) and $0.60 vs. $10.00 for output tokens (12.5x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (220ms faster than GPT-5) with 140 vs. 105 tok/s throughput. GPT-5 leads coding evaluations at 74.9% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 85.4% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs GPT-6 Astra
Qwen 3.8 27B is 98.0% cheaper for input tokens ($0.20 vs. $10.00 per 1M tokens) and $0.60 vs. $50.00 for output tokens (50.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (620ms faster than GPT-6 Astra) with 140 vs. 74 tok/s throughput. GPT-6 Astra leads coding evaluations at 75.7% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 96.0% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs DeepSeek V4 Pro
Qwen 3.8 27B is 84.8% cheaper for input tokens ($0.20 vs. $1.32 per 1M tokens) and $0.60 vs. $3.96 for output tokens (6.6x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (400ms faster than DeepSeek V4 Pro) with 140 vs. 78 tok/s throughput. DeepSeek V4 Pro leads coding evaluations at 80.6% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 90.1% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs DeepSeek R1
Qwen 3.8 27B is 63.6% cheaper for input tokens ($0.20 vs. $0.55 per 1M tokens) and $0.60 vs. $2.19 for output tokens (2.8x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (1330ms faster than DeepSeek R1) with 140 vs. 58 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 49.2% for DeepSeek R1 and 68.0% vs. 71.5% on GPQA Diamond.
Qwen 3.8 27B vs DeepSeek V4 Flash Vision (Exp)
Qwen 3.8 27B is 9.1% cheaper for input tokens ($0.20 vs. $0.22 per 1M tokens) and $0.60 vs. $0.65 for output tokens (1.1x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (190ms faster than DeepSeek V4 Flash Vision (Exp)) with 140 vs. 165 tok/s throughput. DeepSeek V4 Flash Vision (Exp) leads coding evaluations at 77.4% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 91.3% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs OpenAI o3
Qwen 3.8 27B is 90.0% cheaper for input tokens ($0.20 vs. $2.00 per 1M tokens) and $0.60 vs. $8.00 for output tokens (10.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (1030ms faster than OpenAI o3) with 140 vs. 82 tok/s throughput. OpenAI o3 leads coding evaluations at 71.7% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 87.7% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Claude 4 Opus
Qwen 3.8 27B is 98.7% cheaper for input tokens ($0.20 vs. $15.00 per 1M tokens) and $0.60 vs. $75.00 for output tokens (75.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (800ms faster than Claude 4 Opus) with 140 vs. 46 tok/s throughput. Claude 4 Opus leads coding evaluations at 77.2% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 87.1% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Claude 3.7 Sonnet
Qwen 3.8 27B is 93.3% cheaper for input tokens ($0.20 vs. $3.00 per 1M tokens) and $0.60 vs. $15.00 for output tokens (15.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (420ms faster than Claude 3.7 Sonnet) with 140 vs. 72 tok/s throughput. Claude 3.7 Sonnet leads coding evaluations at 70.3% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 78.5% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Gemini 3.1 Pro
Qwen 3.8 27B is 84.0% cheaper for input tokens ($0.20 vs. $1.25 per 1M tokens) and $0.60 vs. $10.00 for output tokens (6.2x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (300ms faster than Gemini 3.1 Pro) with 140 vs. 98 tok/s throughput. Gemini 3.1 Pro leads coding evaluations at 71.8% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 82.4% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Grok 4
Qwen 3.8 27B is 92.0% cheaper for input tokens ($0.20 vs. $2.50 per 1M tokens) and $0.60 vs. $12.50 for output tokens (12.5x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (360ms faster than Grok 4) with 140 vs. 95 tok/s throughput. Grok 4 leads coding evaluations at 73.2% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 86.2% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Llama 4 Maverick
Llama 4 Maverick and Qwen 3.8 27B share identical input pricing at $0.20 per 1M tokens ($0.60 vs. $0.60 output). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (15ms faster than Llama 4 Maverick) with 140 vs. 175 tok/s throughput. Llama 4 Maverick leads coding evaluations at 62.4% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 73.8% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Qwen 3.8 Flash Next
Qwen 3.8 Flash Next is 20.0% cheaper for input tokens ($0.16 vs. $0.20 per 1M tokens) and $0.47 vs. $0.60 for output tokens (1.2x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (160ms faster than Qwen 3.8 Flash Next) with 140 vs. 185 tok/s throughput. Qwen 3.8 Flash Next leads coding evaluations at 81.0% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 91.7% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Gemini 3.5 Flash-Lite
Gemini 3.5 Flash-Lite is 62.5% cheaper for input tokens ($0.07 vs. $0.20 per 1M tokens) and $0.30 vs. $0.60 for output tokens (2.7x cost advantage). In terms of operational performance, Gemini 3.5 Flash-Lite delivers lower response latency at 78ms TTFT (42ms faster than Qwen 3.8 27B) with 265 vs. 140 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 55.4% for Gemini 3.5 Flash-Lite and 68.0% vs. 68.4% on GPQA Diamond.
Qwen 3.8 27B vs OpenAI o4-mini
Qwen 3.8 27B is 81.8% cheaper for input tokens ($0.20 vs. $1.10 per 1M tokens) and $0.60 vs. $4.40 for output tokens (5.5x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (560ms faster than OpenAI o4-mini) with 140 vs. 125 tok/s throughput. OpenAI o4-mini leads coding evaluations at 68.1% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 81.4% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs DeepSeek V3
DeepSeek V3 is 30.0% cheaper for input tokens ($0.14 vs. $0.20 per 1M tokens) and $0.28 vs. $0.60 for output tokens (1.4x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (290ms faster than DeepSeek V3) with 140 vs. 78 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 42.0% for DeepSeek V3 and 68.0% vs. 59.1% on GPQA Diamond.
Qwen 3.8 27B vs Gemma 4 31B IT
Gemma 4 31B IT is 40.0% cheaper for input tokens ($0.12 vs. $0.20 per 1M tokens) and $0.36 vs. $0.60 for output tokens (1.7x cost advantage). In terms of operational performance, Gemma 4 31B IT delivers lower response latency at 115ms TTFT (5ms faster than Qwen 3.8 27B) with 155 vs. 140 tok/s throughput. Gemma 4 31B IT leads coding evaluations at 59.2% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 71.2% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs OpenAI o3-mini
Qwen 3.8 27B is 81.8% cheaper for input tokens ($0.20 vs. $1.10 per 1M tokens) and $0.60 vs. $4.40 for output tokens (5.5x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (830ms faster than OpenAI o3-mini) with 140 vs. 98 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 49.3% for OpenAI o3-mini and 68.0% vs. 79.7% on GPQA Diamond.
Qwen 3.8 27B vs Claude 3.5 Haiku
Qwen 3.8 27B is 75.0% cheaper for input tokens ($0.20 vs. $0.80 per 1M tokens) and $0.60 vs. $4.00 for output tokens (4.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (70ms faster than Claude 3.5 Haiku) with 140 vs. 130 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 40.6% for Claude 3.5 Haiku and 68.0% vs. 41.6% on GPQA Diamond.
Qwen 3.8 27B vs Claude 3.5 Sonnet
Qwen 3.8 27B is 93.3% cheaper for input tokens ($0.20 vs. $3.00 per 1M tokens) and $0.60 vs. $15.00 for output tokens (15.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (190ms faster than Claude 3.5 Sonnet) with 140 vs. 80 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 49.0% for Claude 3.5 Sonnet and 68.0% vs. 59.4% on GPQA Diamond.
Qwen 3.8 27B vs Gemini 2.0 Flash
Gemini 2.0 Flash is 50.0% cheaper for input tokens ($0.10 vs. $0.20 per 1M tokens) and $0.40 vs. $0.60 for output tokens (2.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (160ms faster than Gemini 2.0 Flash) with 140 vs. 185 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 51.8% for Gemini 2.0 Flash and 68.0% vs. 63.6% on GPQA Diamond.
Qwen 3.8 27B vs Gemini 2.5 Flash
Qwen 3.8 27B is 33.3% cheaper for input tokens ($0.20 vs. $0.30 per 1M tokens) and $0.60 vs. $2.50 for output tokens (1.5x cost advantage). In terms of operational performance, Gemini 2.5 Flash delivers lower response latency at 110ms TTFT (10ms faster than Qwen 3.8 27B) with 180 vs. 140 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 58.0% for Gemini 2.5 Flash and 68.0% vs. 70.5% on GPQA Diamond.
Qwen 3.8 27B vs Gemini 2.5 Pro
Qwen 3.8 27B is 84.0% cheaper for input tokens ($0.20 vs. $1.25 per 1M tokens) and $0.60 vs. $10.00 for output tokens (6.2x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (580ms faster than Gemini 2.5 Pro) with 140 vs. 80 tok/s throughput. Gemini 2.5 Pro leads coding evaluations at 63.8% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 74.0% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs GLM-5.3-Flash
GLM-5.3-Flash is 40.0% cheaper for input tokens ($0.12 vs. $0.20 per 1M tokens) and $0.40 vs. $0.60 for output tokens (1.7x cost advantage). In terms of operational performance, GLM-5.3-Flash delivers lower response latency at 90ms TTFT (30ms faster than Qwen 3.8 27B) with 175 vs. 140 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 52.8% for GLM-5.3-Flash and 68.0% vs. 63.5% on GPQA Diamond.
Qwen 3.8 27B vs GLM-5.3
Qwen 3.8 27B is 42.9% cheaper for input tokens ($0.20 vs. $0.35 per 1M tokens) and $0.60 vs. $1.00 for output tokens (1.7x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (40ms faster than GLM-5.3) with 140 vs. 120 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 56.0% for GLM-5.3 and 68.0% vs. 67.0% on GPQA Diamond.
Qwen 3.8 27B vs GPT-4o Mini
GPT-4o Mini is 25.0% cheaper for input tokens ($0.15 vs. $0.20 per 1M tokens) and $0.60 vs. $0.60 for output tokens (1.3x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (100ms faster than GPT-4o Mini) with 140 vs. 140 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 41.0% for GPT-4o Mini and 68.0% vs. 40.2% on GPQA Diamond.
Qwen 3.8 27B vs GPT-4o
Qwen 3.8 27B is 92.0% cheaper for input tokens ($0.20 vs. $2.50 per 1M tokens) and $0.60 vs. $10.00 for output tokens (12.5x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (160ms faster than GPT-4o) with 140 vs. 100 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 48.0% for GPT-4o and 68.0% vs. 53.6% on GPQA Diamond.
Qwen 3.8 27B vs GPT-6.1 Sol
Qwen 3.8 27B is 90.0% cheaper for input tokens ($0.20 vs. $2.00 per 1M tokens) and $0.60 vs. $10.00 for output tokens (10.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (400ms faster than GPT-6.1 Sol) with 140 vs. 135 tok/s throughput. GPT-6.1 Sol leads coding evaluations at 86.4% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 92.4% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Grok 3 Mini
Qwen 3.8 27B is 33.3% cheaper for input tokens ($0.20 vs. $0.30 per 1M tokens) and $0.60 vs. $1.20 for output tokens (1.5x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (30ms faster than Grok 3 Mini) with 140 vs. 160 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 54.0% for Grok 3 Mini and 68.0% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Grok 3
Qwen 3.8 27B is 93.3% cheaper for input tokens ($0.20 vs. $3.00 per 1M tokens) and $0.60 vs. $15.00 for output tokens (15.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (500ms faster than Grok 3) with 140 vs. 88 tok/s throughput. Grok 3 leads coding evaluations at 65.8% SWE-bench Verified vs. 58.4% for Qwen 3.8 27B and 84.6% vs. 68.0% on GPQA Diamond.
Qwen 3.8 27B vs Llama 3.3 70B Instruct
Llama 3.3 70B Instruct is 40.0% cheaper for input tokens ($0.12 vs. $0.20 per 1M tokens) and $0.30 vs. $0.60 for output tokens (1.7x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (25ms faster than Llama 3.3 70B Instruct) with 140 vs. 165 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 42.5% for Llama 3.3 70B Instruct and 68.0% vs. 55.6% on GPQA Diamond.
Qwen 3.8 27B vs Mistral Large 2 (2411)
Qwen 3.8 27B is 90.0% cheaper for input tokens ($0.20 vs. $2.00 per 1M tokens) and $0.60 vs. $6.00 for output tokens (10.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (260ms faster than Mistral Large 2 (2411)) with 140 vs. 75 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 45.2% for Mistral Large 2 (2411) and 68.0% vs. 54.2% on GPQA Diamond.
Qwen 3.8 27B vs o1
Qwen 3.8 27B is 98.7% cheaper for input tokens ($0.20 vs. $15.00 per 1M tokens) and $0.60 vs. $60.00 for output tokens (75.0x cost advantage). In terms of operational performance, Qwen 3.8 27B delivers lower response latency at 120ms TTFT (2480ms faster than o1) with 140 vs. 48 tok/s throughput. Qwen 3.8 27B leads coding evaluations at 58.4% SWE-bench Verified vs. 48.9% for o1 and 68.0% vs. 78.0% on GPQA Diamond.
Frequently Asked Questions & Query Fan-Out
How much does Qwen 3.8 27B cost per 1M tokens including prompt caching?
Qwen 3.8 27B costs $0.20 per million input tokens and $0.60 per million output tokens. Cached input reads cost $0.050 per 1M tokens, reducing repeated context costs by 75%.
What are Qwen 3.8 27B's coding and reasoning benchmark scores?
Qwen 3.8 27B achieves 58.4% on SWE-bench Verified, 91% on HumanEval coding, 68% on GPQA Diamond graduate-level science, 81.2% on MMLU-Pro, and 86.4% on MATH-500, earning an overall Inference Value Ratio (IVR) of 74.9/100.
What is the context window, TTFT latency, and token throughput of Qwen 3.8 27B?
Qwen 3.8 27B supports a 131,072 token context window (16,384 max output tokens) and operates at 120ms average Time to First Token (TTFT) with 140 tokens/sec output streaming throughput.
What are the primary production use cases for Qwen 3.8 27B?
High-performance 27B open weights optimized for local deployment, code generation, and low-latency production pipelines.