~ / local / coding

Running a coding model on your own hardware

Every card and desktop that can hold an open-weight coding model, what it actually runs, how fast, which closed model that matches, and what the hardware costs today. The awkward part is stated up front rather than buried at the bottom.

1Local models with a coding score

24 GBEntry card

~$1,250Cheapest way in

42.2Best local DeepSWE

59.0Lowest Opus you cannot reach

The one thing to know before spending money. Plenty of models are open-weight and plenty of them run on a home machine, but only Qwen3.8-27B has a score on the same coding benchmark the closed models are ranked on. The others that fit on one card — Muse Glimmer 30B, Devstral Small 2, Codestral, StarCoder 2 — are not low-scoring; nobody has tested them on it at all. That is a gap in the evidence, not a verdict on the models, and this page will not pretend otherwise.

What the local ceiling equals in closed models

All figures here are DeepSWE pass@1 from the same table, so they can be compared directly.

Model	DeepSWE	Where it runs	Evidence
Claude Opus 4.8	59.0	API only	measured
Claude Sonnet 5	53.8	API only	measured
GPT-5.4	51.8	API only	measured
Qwen3.8-27B — the local one	42.2	one 24 GB card	Alibaba model card
Claude Sonnet 4.6	29.9	API only	measured

Read the middle row as the answer: the best model you can run at home lands between Claude Sonnet 4.6 and GPT-5.4, and reaches no version of Opus. It is roughly 63% of GPT-5.5 and 72% of Opus 4.8. The Qwen figure is published by Alibaba and has not been independently reproduced; the four closed models are measured entries on the 7 August 2026 DeepSWE table.

Every hardware tier, what it runs, what it costs

Sorted by what the hardware can hold, cheapest route first.

Hardware	Memory / VRAM	Runs Qwen3.8-27B?	Throughput	Matches	Price	Official page
One consumer graphics card
RTX 3090 used	24 GB	Yes, Q4_K_M	not published	Sonnet 4.6 to GPT-5.4	~$1,250	nvidia.com
RTX 4090	24 GB	Yes, Q4_K_M	35–55 tok/s	Sonnet 4.6 to GPT-5.4	$4,399–$4,829	nvidia.com
RTX 5090	32 GB	Yes, up to Q6	60–90 tok/s	Sonnet 4.6 to GPT-5.4	$4,829 new ~$3,590 used	nvidia.com
128 GB unified-memory desktop — more capacity, less speed
AMD Ryzen AI Max+ 395 Strix Halo	128 GB unified ~96 GB to the GPU on Windows, ~108 GB on Linux	Yes, easily	256 GB/s bandwidth	Sonnet 4.6 to GPT-5.4	$3,499 cheapest verified 128GB mini PC	amd.com
NVIDIA DGX Spark GB10	128 GB unified	Yes, easily	273 GB/s bandwidth	Sonnet 4.6 to GPT-5.4	$4,699	nvidia.com
Server and cluster — open weights you cannot run at home
4x H100 80GB	~320 GB	Wrong tool; this tier exists for GLM-5.2	not published	GLM-5.2 scores 43.8	not retail	nvidia.com
Multi-node, 64+ accelerators	~1.4 TB	This tier exists for Kimi K3	not published	Kimi K3 scores 68.5	data-centre scale	huggingface.co

Prices checked 18 August 2026. They are a snapshot, not a live feed, and this class of hardware has been moving quickly. Throughput figures are community measurements at 4-bit and vary with context length, quantisation and speculative decoding, so read them as ranges. Where a figure has not been published we say so rather than invent one.

Check the memory before you check the price. Strix Halo mini PCs are advertised from around $2,000, and that number buys the 64 GB version. The 128 GB configuration that makes this tier interesting is $3,499 at the manufacturer, rising to $3,649 with a larger SSD; AMD's own dev-kit SKU is $3,999. If your goal is only to run a 27B coding model, note that it needs about 17 GB, so the 64 GB machine at $1,999 is already more than enough and a 24 GB graphics card is cheaper still. The 128 GB tier earns its price only when you want to hold models a single card cannot.

The counter-intuitive buy. A used RTX 3090 costs roughly a quarter of a new RTX 4090 and holds exactly the same 24 GB, which is the number that decides whether a model loads at all. The 4090 is not expensive because it is fast — it is expensive because 24 GB of memory is expensive right now. If the goal is running a coding model rather than gaming, the cheapest working setup on this page is the oldest card on it.

The open models, and which ones you can actually run

Open weights and runnable at home are two different questions. This is where most of the confusion lives.

Model	Parameters	VRAM at 4-bit	Coding score	Licence	Runs on
Qwen3.8-27B	28B dense	14–17 GB	DeepSWE 42.2	Apache 2.0	one 24 GB card
Devstral Small 2	24B	~16 GB	not tested on DeepSWE	Apache 2.0	one 24 GB card
Muse Glimmer 30B	29.6B	under 20 GB	not tested on DeepSWE	Apache 2.0	one 24 GB card
Codestral 22B	22B	~14 GB	not tested on DeepSWE	Apache 2.0	one 24 GB card
StarCoder 2 15B	15B	~30 GB	not tested on DeepSWE	OpenRAIL	one 32 GB card
DeepSeek V4 Flash	284B, 13B active	server class pooled, not one card	DeepSWE 53.3	MIT	server
GLM-5.2	753B, 40B active	~370 GB	DeepSWE 43.8	MIT	4x H100
DeepSeek V4 Pro	1.6T, 49B active	~430 GB at INT4 ~862 GB at FP8	DeepSWE 62.8	MIT	16+ H100 for production throughput
Kimi K3	2.8T, 104B active	~1.4 TB	DeepSWE 68.5	bespoke licence	64+ accelerators

Why a mixture-of-experts model still needs all that memory. DeepSeek V4 Pro activates about 49 billion of its 1.6 trillion parameters for any given token, and people reasonably assume that means it only needs memory for 49 billion. It does not. Which experts fire changes from token to token, so every expert has to be resident and ready. The activation count governs how much compute each token costs, not how much memory the model occupies. That is why a model advertised as "49B active" still needs roughly 430 GB at 4-bit, and the same applies to GLM-5.2 and Kimi K3. It is the most common misunderstanding about open-weight models, and it is why an MIT licence changes nothing about whether you can run one.

How much quality quantisation actually costs

Shrinking the model is what makes any of this fit on a consumer card. Here is the published damage.

Precision	Quality retained	Memory / VRAM	Speed	Use it when
FP16	reference	100%	baseline	You have memory to spare
Q8_0	effectively lossless within 0.02–0.05 perplexity of FP16	~50%	slightly faster	Quality-critical work and evaluation
Q4_K_M	95–97% of FP16	~35%	2–3x faster	The default for a 24 GB card
Q4 generally	perplexity 1–3% worse	~25%	3.5–3.8x faster	When memory is the binding constraint

One directly relevant measurement: on HumanEval a 32B model scored identically at Q4_K_M, AWQ and bitsandbytes 4-bit, which suggests 4-bit does not meaningfully damage coding ability at this size. Perplexity is a proxy rather than a coding score, so treat the percentages as the shape of the trade-off, not as points on the DeepSWE scale.

Starting Qwen3.8-27B

Shortest path first.

ollama pull qwen3.8:27b
ollama run qwen3.8:27b

To drive it with a real agent harness rather than a chat box, Ollama exposes an Anthropic-compatible endpoint, so Claude Code can point straight at it:

export ANTHROPIC_BASE_URL=http://localhost:11434
export ANTHROPIC_AUTH_TOKEN=ollama
claude

Remap ANTHROPIC_MODEL and ANTHROPIC_SMALL_FAST_MODEL to the local model name. For a vLLM server or any OpenAI-format endpoint, put LiteLLM in front as a translating gateway. Note that no published benchmark covers Claude Code driving a local model, so the harness scores elsewhere on this site do not transfer to that combination.

Full Qwen3.8-27B setup and score comparison Hardware table on the leaderboard All local deployment guides