Locally Explains

Claude Grade Coding On A Cheap 12GB GPU (Local AI)

Every figure in the video, with where it was published. Checked 2026-10-10.

Qwen3.8-27B is a 27 billion parameter open model from Alibaba’s Qwen team. Native context 262,144 tokens, extensible to 1,000,000. Source: Qwen3.8-27B model card, https://huggingface.co/Qwen/Qwen3.8-27B
From the same card’s benchmark table (Qwen3.8-27B vs Opus4.6 Max):
- SWE-bench Pro: 61.7 vs 53.4 (Qwen ahead by 8.3)
- LiveCodeBench v6: 90.3 vs 88.8 (Qwen ahead by 1.5)
- Terminal Bench 2.1 (Terminus): 73.0 vs 78.2 (Claude ahead by 5.2)
The card’s SWE-bench footnote: all models are evaluated with the Claude Code harness at temperature 1.0, top_p 0.95 and a 256K context window, except Opus4.6 Max, which uses its officially reported score. Source: https://huggingface.co/Qwen/Qwen3.8-27B
Claude Opus 4.8 scores 69.2 on SWE-bench Pro (Anthropic system card; also reported by third-party summaries). Sources: https://www.anthropic.com/news/claude-opus-4-8, https://vellum.ai/blog/claude-opus-4-8-benchmarks-explained

ISTA-DASLab’s GSQ-RCO GGUF builds of Qwen3.8-27B assign a quantization type to every tensor, chosen by per-tensor sensitivity.
IQ2_XS: 2.50 bits per weight, 8.4 GB, LiveCodeBench v6 76.57.
BF16 baseline on the same table: 53.8 GB, LiveCodeBench v6 85.71. Difference: 9.14 points.
IQ3_S: 3.50 bits per weight, 11.8 GB, LiveCodeBench v6 85.71 (the recommended build).
Unsloth UD-IQ2_S at the same 8.4 GB: LiveCodeBench v6 72.00.
The optional MTP builds are about 0.35 GB larger (they carry the multi-token prediction head). Source: https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF

The official Aider Polyglot leaderboard lists no current Qwen3.8 entry and no matched Claude row for this model. Source: https://aider.chat/docs/leaderboards/
An independent run in a public benchmark repository: Qwen3.8-27B on aider-polyglot, thinking off, 32k slot: pass@1 25.8%, pass@2 68.9%, 225 of 225 tasks, 2 malformed responses, median 12.3 minutes per task. The page has no Claude result. Source: https://github.com/ccarokh/AI-Model-Benchmark-Runs/blob/main/use-cases/coding.md

RTX 4060 8 GB: Unsloth IQ4_XS, about 14 GB, roughly 7 GB on the card (26 layers) and the rest in system RAM; 70,000 token context target; about 4.2 to 5 tokens per second, falling as the context fills. Source: https://brainwagon.org/blog/2026-08-24-running-the-qwen-3-8-27b-dense-model-on-8gb-vram
RTX 3060 12 GB: about 10 tok/s with IQ3_XXS; 8 to 9 tok/s with Q4_K_M, falling to 4 or 5 past 30K context (community thread). Source: https://www.reddit.com/r/LocalLLM/comments/1wldram/qwen38_27b_what_is_realistic_tokens_per_second/
RTX 3060 12 GB, patched build with multi-token prediction: almost 30 tok/s on a fixed C++ task at 64K context. Source: https://www.reddit.com/r/unsloth/comments/1w0cn86/qwen3827b_30_toks_at_64k_on_one_rtx_3060_12_gb/
RTX 5060 Ti 16 GB: around 40 to 50 tok/s at 32K context with a 3 bit quant and MTP; another owner a little over 20 tok/s on a similar size quant with a different context and runtime. Source: https://www.reddit.com/r/LocalLLM/comments/1wldram/qwen38_27b_what_is_realistic_tokens_per_second/

Claude Pro $20 a month; Claude Max from $100 a month. https://claude.com/pricing
Used RTX 3060 12 GB: RigPrice going rate (median ask) $370, 62 active listings; the cheapest screened pick was $340 in the October 10 morning snapshot and $350 later that day. https://rigprice.com/gpu/rtx-3060-12gb/
GIGABYTE RTX 5060 8 GB, new: $459.99 at Newegg, the lowest new RTX 5060 sorted by price. https://www.newegg.com/p/pl?d=rtx+5060+8gb&N=100007709&Order=1
RTX 5060 Ti 16 GB, new: $789 in PC Gamer’s October 8 price watch; $799.99 lowest new on Newegg on October 10. https://www.pcgamer.com/hardware/graphics-cards/graphics-card-price-watch-deals/, https://www.newegg.com/p/pl?d=rtx+5060+ti+16gb&N=100007709&Order=1

Card price divided by plan price, before power, the rest of the computer and the value of a hosted service:

card	price	at $20 a month	at $100 a month
GIGABYTE RTX 5060 8 GB	$460	23 months	4.6 months
used RTX 3060 12 GB, going ask	$370	18.5 months	3.7 months
used RTX 3060 12 GB, screened pick	$340	17 months	3.4 months
RTX 5060 Ti 16 GB	$789	about 40 months	about 8 months

The Reddit owner reports (RTX 3060 and RTX 5060 Ti speeds) could not be opened from our side on the day; they are quoted as the script’s writer cited them.
Opus 4.8’s 69.2 is taken from summaries of Anthropic’s system card; the PDF itself was too large to read here.