What is the best local LLM for coding on 8 GB VRAM?

Qwen3.5 9B: The best coding model for 8–12 GB cards. At Q4_K_M it takes 6.3 GB of 7.0 GB usable with an 8K context, about 30 tokens per second estimated on a GeForce RTX 4060.

What is the best local LLM for coding on 12 GB VRAM?

Qwen3.5 9B: The best coding model for 8–12 GB cards. At Q4_K_M it takes 6.3 GB of 10.6 GB usable with an 8K context, about 40 tokens per second estimated on a GeForce RTX 3060 12 GB.

What is the best local LLM for coding on 16 GB VRAM?

Qwen3.5 9B: The best coding model for 8–12 GB cards. At Q4_K_M it takes 6.3 GB of 14.4 GB usable with an 8K context, about 107 tokens per second estimated on a GeForce RTX 5080.

What is the best local LLM for coding on 24 GB VRAM?

Qwen3.8 27B: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. At Q4_K_M it takes 16.7 GB of 22.4 GB usable with an 8K context, about 39 tokens per second estimated on a GeForce RTX 4090.

What is the best local LLM for coding on 32 GB VRAM?

Qwen3.8 27B: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. At Q4_K_M it takes 16.7 GB of 30.4 GB usable with an 8K context, about 69 tokens per second estimated on a GeForce RTX 5090.

What is the best local LLM for coding on 48 GB VRAM?

Qwen3.8 27B: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. At Q4_K_M it takes 16.7 GB of 46.4 GB usable with an 8K context, about 37 tokens per second estimated on a RTX 6000 Ada.

What is the best local LLM for coding on 64 GB unified memory?

Qwen3.8 27B: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. At Q4_K_M it takes 16.7 GB of 48.0 GB usable with an 8K context, about 15 tokens per second estimated on a M4 Max · 64 GB.

What is the best local LLM for coding on 128 GB unified memory?

Ling 3.0 Flash 124B-A5B: SWE-bench Pro 56.6 claimed at 5B active — the 128 GB-class coding pick if you cannot fit DeepSeek V4. At Q4_K_M it takes 70.4 GB of 96.0 GB usable with an 8K context, about 50 tokens per second estimated on a M4 Max · 128 GB.

Best local LLM for coding · 2026

The best local LLM for coding, ranked for your VRAM

Updated 8 Oct 2026 1 new this month

The coding shortlist is ordered by what the vendors' own evaluations and the local-model community report for software-engineering agents, code editing and autocomplete — SWE-bench Verified, Terminal-Bench, DeepSWE — as of the catalogue date. For each VRAM class below we list every model on that list that actually fits at Q4_K_M with an 8K context, so the top row is the pick and the rest are what you give up by choosing it. Numbers are computed, not benchmarked; how the shortlist works.

A coding agent needs context more than most jobs. The "max context" column is the longest window the model can hold on this class at that quantisation — if it reads under 32K, expect to drop a quantisation step or run the model with a q8_0 KV cache before pointing an agent at a whole repository.

Best coding LLM for 8 GB VRAM

7.0 GB usable · speeds for a GeForce RTX 4060

Qwen3.5 9B is the pick: The best coding model for 8–12 GB cards. 4 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	31K	30	Tight fit
2	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	35K	31	Tight fit
3	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	120K	63	Runs great
4	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	38K	38	Runs great
5	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	129	Runs great

Best coding LLM for 12 GB VRAM

10.6 GB usable · speeds for a GeForce RTX 3060 12 GB

Qwen3.5 9B is the pick: The best coding model for 8–12 GB cards. 4 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	146K	40	Runs great
2	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	150K	41	Runs great
3	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	236K	83	Runs great
4	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	104K	51	Runs great
5	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	171	Runs great

Best coding LLM for 16 GB VRAM

14.4 GB usable · speeds for a GeForce RTX 5080

Qwen3.5 9B is the pick: The best coding model for 8–12 GB cards. 4 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	256K	107	Runs great
2	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	256K	110	Runs great
3	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	256K	222	Runs great
4	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	128K	136	Runs great
5	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	455	Runs great

Best coding LLM for 24 GB VRAM

22.4 GB usable · speeds for a GeForce RTX 4090

Qwen3.8 27B is the pick: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. 11 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Qwen3.8 27B Alibaba · Apache 2.0	Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB.	27.8B	Q4_K_M · 16.7 GB	98K	39	Runs great
2	Qwen3.6 27B Alibaba · Apache 2.0	77.2 SWE-bench Verified at 27B; the spring-2026 local coding standard.	27.8B	Q4_K_M · 16.7 GB	98K	39	Runs great
3	Qwen3.6 35B-A3B Alibaba · Apache 2.0	Agentic coding at 3B-active speed — the best MoE under 40B for it.	35.9B MoE	Q4_K_M · 20.9 GB	82K	126	Tight fit
4	Devstral Small 2 24B Mistral AI · Apache 2.0	Tuned for software-engineering agents like OpenHands and Cline.	24B	Q4_K_M · 15.3 GB	53K	45	Runs great
5	GLM-4.7-Flash 30B-A3B Z.ai · MIT	MIT-licensed 30B-A3B built for agentic coding; 60–80 tok/s reported on a 4090.	31.2B MoE	Q4_K_M · 18.6 GB	82K	139	Runs great
6	Ornith 1.5 35B-A3B Ornith AI · MIT	Reasoning-first coding build on the Qwen3.6 MoE; thinks before it edits.	35.9B MoE	Q4_K_M · 20.9 GB	82K	126	Tight fit
7	Qwen3 Coder 30B-A3B Alibaba · Apache 2.0	The most-downloaded local code model; 256K window.	30.5B MoE	Q4_K_M · 18.5 GB	49K	126	Runs great
8	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	256K	112	Runs great
9	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	256K	115	Runs great
10	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	256K	233	Runs great
11	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	128K	142	Runs great
12	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	478	Runs great

Best coding LLM for 32 GB VRAM

30.4 GB usable · speeds for a GeForce RTX 5090

Qwen3.8 27B is the pick: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. 11 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Qwen3.8 27B Alibaba · Apache 2.0	Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB.	27.8B	Q4_K_M · 16.7 GB	226K	69	Runs great
2	Qwen3.6 27B Alibaba · Apache 2.0	77.2 SWE-bench Verified at 27B; the spring-2026 local coding standard.	27.8B	Q4_K_M · 16.7 GB	226K	69	Runs great
3	Qwen3.6 35B-A3B Alibaba · Apache 2.0	Agentic coding at 3B-active speed — the best MoE under 40B for it.	35.9B MoE	Q4_K_M · 20.9 GB	256K	225	Runs great
4	Devstral Small 2 24B Mistral AI · Apache 2.0	Tuned for software-engineering agents like OpenHands and Cline.	24B	Q4_K_M · 15.3 GB	104K	80	Runs great
5	GLM-4.7-Flash 30B-A3B Z.ai · MIT	MIT-licensed 30B-A3B built for agentic coding; 60–80 tok/s reported on a 4090.	31.2B MoE	Q4_K_M · 18.6 GB	198K	247	Runs great
6	Ornith 1.5 35B-A3B Ornith AI · MIT	Reasoning-first coding build on the Qwen3.6 MoE; thinks before it edits.	35.9B MoE	Q4_K_M · 20.9 GB	256K	225	Runs great
7	Qwen3 Coder 30B-A3B Alibaba · Apache 2.0	The most-downloaded local code model; 256K window.	30.5B MoE	Q4_K_M · 18.5 GB	134K	225	Runs great
8	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	256K	200	Runs great
9	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	256K	205	Runs great
10	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	256K	414	Runs great
11	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	128K	253	Runs great
12	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	850	Runs great

Best coding LLM for 48 GB VRAM

46.4 GB usable · speeds for a RTX 6000 Ada

Qwen3.8 27B is the pick: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. 11 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Qwen3.8 27B Alibaba · Apache 2.0	Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB.	27.8B	Q4_K_M · 16.7 GB	256K	37	Runs great
2	Qwen3.6 27B Alibaba · Apache 2.0	77.2 SWE-bench Verified at 27B; the spring-2026 local coding standard.	27.8B	Q4_K_M · 16.7 GB	256K	37	Runs great
3	Qwen3.6 35B-A3B Alibaba · Apache 2.0	Agentic coding at 3B-active speed — the best MoE under 40B for it.	35.9B MoE	Q4_K_M · 20.9 GB	256K	120	Runs great
4	Devstral Small 2 24B Mistral AI · Apache 2.0	Tuned for software-engineering agents like OpenHands and Cline.	24B	Q4_K_M · 15.3 GB	206K	43	Runs great
5	GLM-4.7-Flash 30B-A3B Z.ai · MIT	MIT-licensed 30B-A3B built for agentic coding; 60–80 tok/s reported on a 4090.	31.2B MoE	Q4_K_M · 18.6 GB	198K	133	Runs great
6	Ornith 1.5 35B-A3B Ornith AI · MIT	Reasoning-first coding build on the Qwen3.6 MoE; thinks before it edits.	35.9B MoE	Q4_K_M · 20.9 GB	256K	120	Runs great
7	Qwen3 Coder 30B-A3B Alibaba · Apache 2.0	The most-downloaded local code model; 256K window.	30.5B MoE	Q4_K_M · 18.5 GB	256K	120	Runs great
8	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	256K	107	Runs great
9	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	256K	110	Runs great
10	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	256K	222	Runs great
11	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	128K	136	Runs great
12	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	455	Runs great

Best coding LLM for 64 GB unified memory

48.0 GB usable · speeds for a M4 Max · 64 GB

Qwen3.8 27B is the pick: Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB. 11 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Qwen3.8 27B Alibaba · Apache 2.0	Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB.	27.8B	Q4_K_M · 16.7 GB	256K	15	Runs great
2	Qwen3.6 27B Alibaba · Apache 2.0	77.2 SWE-bench Verified at 27B; the spring-2026 local coding standard.	27.8B	Q4_K_M · 16.7 GB	256K	15	Runs great
3	Qwen3.6 35B-A3B Alibaba · Apache 2.0	Agentic coding at 3B-active speed — the best MoE under 40B for it.	35.9B MoE	Q4_K_M · 20.9 GB	256K	58	Runs great
4	Devstral Small 2 24B Mistral AI · Apache 2.0	Tuned for software-engineering agents like OpenHands and Cline.	24B	Q4_K_M · 15.3 GB	216K	17	Runs great
5	GLM-4.7-Flash 30B-A3B Z.ai · MIT	MIT-licensed 30B-A3B built for agentic coding; 60–80 tok/s reported on a 4090.	31.2B MoE	Q4_K_M · 18.6 GB	198K	63	Runs great
6	Ornith 1.5 35B-A3B Ornith AI · MIT	Reasoning-first coding build on the Qwen3.6 MoE; thinks before it edits.	35.9B MoE	Q4_K_M · 20.9 GB	256K	58	Runs great
7	Qwen3 Coder 30B-A3B Alibaba · Apache 2.0	The most-downloaded local code model; 256K window.	30.5B MoE	Q4_K_M · 18.5 GB	256K	58	Runs great
8	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	256K	42	Runs great
9	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	256K	43	Runs great
10	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	256K	87	Runs great
11	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	128K	53	Runs great
12	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	179	Runs great

Best coding LLM for 128 GB unified memory

96.0 GB usable · speeds for a M4 Max · 128 GB

Ling 3.0 Flash 124B-A5B is the pick: SWE-bench Pro 56.6 claimed at 5B active — the 128 GB-class coding pick if you cannot fit DeepSeek V4. 12 other models on the coding shortlist also fit, in order of capability:

#	Model	Why	Params	Quant · total	Max context	Tok/s est.
1	Ling 3.0 Flash 124B-A5B inclusionAI · MIT	SWE-bench Pro 56.6 claimed at 5B active — the 128 GB-class coding pick if you cannot fit DeepSeek V4.	124B MoE	Q4_K_M · 70.4 GB	256K	50	Runs great
2	Qwen3.8 27B Alibaba · Apache 2.0	Terminal-Bench 73, DeepSWE 42 — a generation ahead of anything else that fits 24 GB.	27.8B	Q4_K_M · 16.7 GB	256K	20	Runs great
3	Qwen3.6 27B Alibaba · Apache 2.0	77.2 SWE-bench Verified at 27B; the spring-2026 local coding standard.	27.8B	Q4_K_M · 16.7 GB	256K	20	Runs great
4	Qwen3.6 35B-A3B Alibaba · Apache 2.0	Agentic coding at 3B-active speed — the best MoE under 40B for it.	35.9B MoE	Q4_K_M · 20.9 GB	256K	77	Runs great
5	Devstral Small 2 24B Mistral AI · Apache 2.0	Tuned for software-engineering agents like OpenHands and Cline.	24B	Q4_K_M · 15.3 GB	384K	23	Runs great
6	GLM-4.7-Flash 30B-A3B Z.ai · MIT	MIT-licensed 30B-A3B built for agentic coding; 60–80 tok/s reported on a 4090.	31.2B MoE	Q4_K_M · 18.6 GB	198K	84	Runs great
7	Ornith 1.5 35B-A3B Ornith AI · MIT	Reasoning-first coding build on the Qwen3.6 MoE; thinks before it edits.	35.9B MoE	Q4_K_M · 20.9 GB	256K	77	Runs great
8	Qwen3 Coder 30B-A3B Alibaba · Apache 2.0	The most-downloaded local code model; 256K window.	30.5B MoE	Q4_K_M · 18.5 GB	256K	77	Runs great
9	Qwen3.5 9B Alibaba · Apache 2.0	The best coding model for 8–12 GB cards.	9.65B	Q4_K_M · 6.3 GB	256K	56	Runs great
10	Ornith 1.5 9B Ornith AI · MIT	A coding-agent reasoning build on the Qwen3.5 9B architecture.	9.41B	Q4_K_M · 6.1 GB	256K	58	Runs great
11	Qwen3.5 4B Alibaba · Apache 2.0	A coding agent in 3.4 GB — the 6–8 GB answer.	4.66B	Q4_K_M · 3.5 GB	256K	116	Runs great
12	Qwen2.5-Coder 7B Alibaba · Apache 2.0	Still the best fill-in-the-middle autocomplete model under 8B.	7.62B	Q4_K_M · 5.3 GB	128K	71	Runs great
13	Qwen3.5 2B Alibaba · Apache 2.0	Autocomplete on 4 GB.	2.27B	Q4_K_M · 2.0 GB	256K	239	Runs great

The whole coding shortlist, and what each one needs

Every model on the list, best first, with the smallest discrete card that runs it with 15% headroom at the recommended quantisation and 8K context.

#	Model	Params	Weights	Needs	Ollama tag
1	DeepSeek V4.1 Flash 552B-A16B New Twice V4 Flash's backbone (552B, 16B active per generated token) plus 196B of Engram lookup tables (189 GiB at FP8) that DwarfStar streams from SSD, so they are not counted here. A 512 GB Mac model, and as of October 2026 no mainline llama.cpp or Ollama build: DwarfStar on a Mac, vLLM across four GPUs. DeepSeek puts the cache at 890 bytes/token; modelled conservatively as 4 latent layers plus a 128-token window.	552B MoE	310.4 GB @ Q4_K_M	more than one card	—
2	Kimi K2.6 1T-A32B The open coding-agent benchmark leader of spring 2026 (80.2 SWE-bench). A 512 GB Mac Studio pair, or a ceiling.	1027B MoE	577.5 GB @ Q4_K_M	more than one card	—
3	GLM-5.3 744B-A40B Same base as GLM-5.2, new post-training: Z.ai's most capable open-weights coder (Terminal-Bench 3.0 28.3, DeepSWE 66.9). Not MIT any more — its own GLM-5.3 licence. Same ceiling as 5.2: 512 GB of unified memory at Q4.	753B MoE	423.4 GB @ Q4_K_M	more than one card	—
4	GLM-5.3-Flash 320B-A18B The first natively multimodal GLM-5 and the first hybrid: 34 linear-attention blocks and 11 sparse-attention blocks with a 512-wide latent cache, so a 1M window stays affordable. Z.ai says it beats GLM-5.2 at 18B active; MIT.	321B MoE	180.5 GB @ Q4_K_M	more than one card	—
5	DeepSeek V4 Flash 284B-A13B The V4 for 192–256 GB machines: Q3 squeezes into 192 GB, Q4 wants 256 GB, and 128 GB falls short even at Q2. Cache is modelled as a 576-wide latent; V4 compresses it further at long context, so this is conservative.	284B MoE	159.7 GB @ Q4_K_M	192 GB card	—
6	Qwen3.8-Flash-Next 180B-A6B The open preview of the Qwen4 architecture: a 125B-A6B hybrid (Gated DeltaNet + sparse attention, KV cache on 12 of 48 blocks) plus a 51B n-gram embedding and a 4B draft head — 180B on disk, 6B active. The hosted "Qwen3.8-Flash" is this model with a 1M window.	180B MoE	101.2 GB @ Q4_K_M	141 GB card	qwen3.8-flash-next:125b-a6b-nvfp4
7	Ling 3.0 Flash 124B-A5B A 124B hybrid (5 linear-attention layers per MLA layer) with 5.1B active: SWE-bench Pro 56.6 and AIME 93 claimed. Built for 96–128 GB machines.	124B MoE	69.7 GB @ Q4_K_M	141 GB card	—
8	Qwen3.8 27B The current default local Qwen: dense 27B, text + image + video, 262K context. Only 16 of its 64 blocks keep a KV cache, so long context is cheap.	27.8B	15.6 GB @ Q4_K_M	24 GB card	qwen3.8:27b
9	Qwen3.6 27B The 24 GB coding pick of spring 2026 (77.2 SWE-bench Verified). Same shape as 3.8, one generation behind.	27.8B	15.6 GB @ Q4_K_M	24 GB card	qwen3.6:27b
10	Qwen3.6 35B-A3B Mixture of experts with ~3B active: the fastest serious model a 24 GB card runs, and the best MoE under 40B on agentic coding.	35.9B MoE	20.2 GB @ Q4_K_M	32 GB card	qwen3.6:35b-a3b
11	Devstral Small 2 24B Built for software-engineering agents (OpenHands, Cline). Dense 24B, 384K window. Not in the Ollama library.	24B	13.5 GB @ Q4_K_M	20 GB card	—
12	GLM-4.7-Flash 30B-A3B MIT-licensed 30B-A3B tuned for agentic coding, with a DeepSeek-style latent KV cache. 60–80 tok/s reported on a 4090.	31.2B MoE	17.5 GB @ Q4_K_M	24 GB card	glm-4.7-flash:latest
13	Ornith 1.5 35B-A3B A reasoning-first MIT build on the Qwen3.6 35B-A3B architecture (thinks before every answer). Same VRAM as its base.	35.9B MoE	20.2 GB @ Q4_K_M	32 GB card	ornith-1.5:35b
14	Qwen3 Coder 30B-A3B Agentic coding MoE with a 256K native window. Still the most-downloaded local code model.	30.5B MoE	17.1 GB @ Q4_K_M	24 GB card	qwen3-coder:30b
15	Qwen3.5 9B The default for 8–12 GB cards in 2026: beats every older 8B on every published benchmark, with vision.	9.65B	5.4 GB @ Q4_K_M	10 GB card	qwen3.5:9b
16	Ornith 1.5 9B The small Ornith: a coding-agent reasoning build on the Qwen3.5 9B architecture. Same VRAM as its base, thinks before every answer.	9.41B	5.3 GB @ Q4_K_M	10 GB card	ornith-1.5:9b
17	Qwen3.5 4B The 8 GB coding agent. Q4 lands near 3.4 GB, leaving room for a real context window.	4.66B	2.6 GB @ Q4_K_M	8 GB card	qwen3.5:4b
18	Qwen2.5-Coder 7B The standard local autocomplete model — small enough to keep resident all day, and still the best FIM model under 8B.	7.62B	4.3 GB @ Q4_K_M	8 GB card	qwen2.5-coder:7b
19	Qwen3.5 2B Phone-class, and multimodal. Replaces Llama 3.2 3B as the "it runs on anything" answer.	2.27B	1.3 GB @ Q4_K_M	8 GB card	qwen3.5:2b

Autocomplete and agents want different things. Fill-in-the-middle completion is latency-bound and happy with a 4–9B model; an agent that edits files and runs tests wants the strongest model you can hold with a 32K+ window, and will wait for it. If you are unsure which card to buy for this, the GPU ranking says what each class unlocks.