llmfind llms should run on premises

y coding indexx $ per 1M tokens, 3000 in : 1500 out, logsize output tokens/s ● api● open weights, priced by card write-off◌ new this fortnight

try it — click a dot: the model opens on arena.ai (or OpenRouter chat when arena does not list it) and this prompt lands in your clipboard, paste it there:

Write a python function that parses ISO-8601 durations like P3DT4H30M into seconds. Add doctests.

model		coding	$ / 1M	tok/s	runs on	since
Claude Fable 5.1 anthropic	●	80.7	23.3	53		2026-09	try
GPT-5.6 Sol openai	●	78.3	9.33	106		2026-07	try
Claude Opus 5 anthropic	●	78	11.7	49		2026-07	try
GPT-6 Astra openai	●	76.9	23.3	49		2026-09	try
GPT-5.6 Terra openai	●	76.7	5.33	95		2026-07	try
Claude Fable 5 anthropic	●	76.5	23.3	66		2026-06	try
Gemini 3.8 Flash google	●	76.3	1.93	136		2026-09	try
Qwen3.8 Max (0902) qwen	●	76.2	3.33			2026-09	try
Gemini 3.7 Flash google	●	76.1	1.75			2026-08	try
Grok 4.6 x-ai	●	75.9	3.33	63		2026-08	try
Muse Spark 1.3 meta	●	75.8	2.25	146		2026-09	try
GPT-5.5 openai	●	74.9	13.3	75		2026-04	try
GLM 5.3 z-ai	●	74.8	0.108	72	6× B200 753B, 35B active	2026-08	try
Claude Opus 4.8 anthropic	●	74.3	11.7	58		2026-05	try
Claude Opus 4.7 anthropic	●	73.6	11.7	43		2026-04	try
Grok 4.5 x-ai	●	72.4	3.33	57		2026-07	try
Muse Spark 1.2 meta	●	72.2	2.25	228		2026-08	try
Kimi K3 moonshotai	●	72	5.46	50	api price	2026-07	try
Qwen3.8 2.4T A95B qwen	●	71.9	3.33	40	api price	2026-08	try
GLM 5.3 Flash z-ai	●	71.5	0.0411	55	3× B200 321B, 13B active	2026-08	try
Claude Sonnet 5 anthropic	●	71.5	4.67	57		2026-06	try
GPT-5.6 Luna openai	●	71.4	0.533	105		2026-07	try
Muse Spark 1.1 meta	●	71.3	2.25	224		2026-07	try
GPT-5.4 openai	●	71.1	6.67	56		2026-03	try
Gemini 3.5 Flash google	●	70.1	4	191		2026-05	try
Gemini 3.6 Flash google	●	69.2	1.75	200		2026-07	try
DeepSeek V4 Flash 0731 deepseek	●	69.1	0.0328	123	3× B200 304B, 11B active	2026-07	try
DeepSeek V4 Pro 0813 deepseek	●	68.8	1.63	79	api price	2026-08	try
GLM 5.2 z-ai	●	68.8	0.108	193	6× B200 753B, 35B active	2026-06	try
Gemini 3.1 Pro Preview google	●	68.8	5.33	134		2026-02	try
Qwen3.8 27B qwen	●	68.1	0.0853	36	1× B200 28B	2026-08	try
Qwen3.7 Max qwen	●	66	2.46	200		2026-05	try
DeepSeek V4 Flash Vision Exp deepseek	●	65	0.0329		3× B200 305B, 11B active	2026-08	try
Claude Sonnet 4.6 anthropic	●	63	7	44		2026-02	try
Kimi K2.6 moonshotai	●	61.8	0.0985	42	7× B200 1027B, 32B active	2026-04	try
Kimi K2.7 Code moonshotai	●	60.8	0.0985	57	7× B200 1027B, 32B active	2026-06	try
MiMo-V2.5-Pro xiaomi	●	60.2	0.0982	52	7× B200 1023B, 32B active	2026-04	try
Hy3 preview tencent	●	58.8	0.0573		2× B200 299B, 19B active	2026-04	try
MiniMax M3 minimax	●	58.6	0.0615	121	3× B200 427B, 20B active	2026-05	try
Ling 3.0 Flash VL inclusionai	●	57	0.00899		1× B200 125B, 3B active	2026-09	try
MiMo-V2.5 xiaomi	●	56.8	0.0447	79	3× B200 311B, 15B active	2026-04	try
DeepSeek V4 Flash 0423 deepseek	●	56.2	0.0314		2× B200 291B, 10B active	2026-04	try
GPT-5.4 Nano openai	●	56.1	0.55	169		2026-03	try
GPT-5.4 Mini openai	●	56.1	2	204		2026-03	try
Qwen3.7 Plus qwen	●	55.9	0.64	56		2026-06	try
GLM 5.1 z-ai	●	55.8	0.109	77	6× B200 754B, 35B active	2026-04	try
Qwen3.6 Plus qwen	●	54.5	0.867	56		2026-04	try
Qwen3.6 27B qwen	●	53.7	0.0853	56	1× B200 28B	2026-04	try
Inkling Small thinkingmachines	●	52.9	0.0287	81	2× B200 266B, 9B active	2026-07	try
Solar Pro 4 upstage	●	52.7	0.3	58		2026-08	try
MiniMax M2.7 minimax	●	52.6	0.0329	49	2× B200 229B, 11B active	2026-03	try
Inkling thinkingmachines	●	52.1	0.103	139	7× B200 952B, 33B active	2026-07	try
Claude Sonnet 4.5 anthropic	●	52.1	7			2025-09	try
Grok Build 0.1 x-ai	●	51.5	1.33			2026-05	try
Ling 3.0 Flash inclusionai	●	50.6	0.00918	309	1× B200 127B, 3B active	2026-07	try
GPT-5.1 openai	●	49.4	4.17	95		2025-11	try
Gemini 3.5 Flash Lite google	●	49.3	1.03	392		2026-07	try
Nemotron 3 Ultra nvidia	●	49.3	0.169		4× B200 550B, 55B active	2026-06	try
Muse Glimmer 30B meta	●	49	0.0914		1× B200 30B	2026-08	try
Qwen3.5 397B A17B qwen	●	48.2	0.0522	59	3× B200 397B, 17B active	2026-02	try
Mistral Medium 3.5 mistralai	●	46.9	3.5	172		2026-04	try
Kimi K2.5 moonshotai	●	46.8	0.0985	52	7× B200 1027B, 32B active	2026-01	try
GLM 4.6 z-ai	●	45.8	0.0822	38	3× B200 357B, 27B active	2025-09	try
Qwen3.5-122B-A10B qwen	●	45.7	0.0307	143	1× B200 122B, 10B active	2026-02	try
LongCat 2.0 meituan	●	45.3	0.6	60	api price	2026-07	try
GLM 4.7 z-ai	●	45.3	0.0825	39	3× B200 358B, 27B active	2025-12	try
DeepSeek V3.2 deepseek	●	44.2	0.0987		5× B200 685B, 32B active	2025-12	try
Claude Haiku 4.5 anthropic	●	43.9	2.33			2025-10	try
DeepSeek V3.1 Terminus deepseek	●	43.5	0.0985		5× B200 685B, 32B active	2025-09	try
Gemma 4 31B google	●	43.4	0.096	35	1× B200 31B	2026-04	try
Grok 4.3 x-ai	●	42.2	1.67	72		2026-04	try
Qwen3.6 35B A3B qwen	●	41.9	0.00921	126	1× B200 35B, 3B active	2026-04	try
Step 3.7 Flash stepfun	●	39.6	0.0258	176	2× B200 201B, 8B active	2026-05	try
Gemma 4 26B A4B google	●	39.3	0.0123	87	1× B200 26B, 4B active	2026-04	try
GPT-5 openai	●	37.8	4.17	91		2025-08	try
Nemotron 3 Super nvidia	●	37.7	0.0369		1× B200 120B, 12B active	2026-03	try
Qwen3.5-35B-A3B qwen	●	37	0.00921	148	1× B200 35B, 3B active	2026-02	try
North Mini Code cohere	●	36.5	0.00878	99	1× B200 30B, 3B active	2026-06	try
Qwen3 Coder Next qwen	●	36.2	0.00717	78	1× B200 80B, 2B active	2026-02	try
Gemini 3.1 Flash Lite Preview google	●	34.7	0.667			2026-03	try
Gemini 2.5 Pro google	●	33.3	4.17	120		2025-06	try
Devstral 2 2512 mistralai	●	31.3	0.384		1× B200 125B	2025-12	try
Mercury 2 inception	●	31.1	0.417	512		2026-03	try
gpt-oss-120b openai	●	30.4	0.0168	217	1× B200 117B, 5B active	2025-08	try
Qwen3.5-9B qwen	●	28.7	0.0296	153	1× B200 10B	2026-03	try
Command A+ cohere	●	27.8	0.7	247		2026-09	try
Nemotron 3.5 Lightning nvidia	●	26.8	0.00921	293	1× B200 30B, 3B active	2026-08	try
Mistral Small 4 mistralai	●	26.6	0.0172	160	1× B200 119B, 6B active	2026-03	try
Ling 3.0 Tiny inclusionai	●	26.5	0.00227	63	1× B200 8B, 1B active	2026-08	try
GPT-4o (2024-05-13) openai	●	24.2	8.33			2024-05	try
Nova 2 Lite amazon	●	23	1.03			2025-12	try
Granite 4.2 8B ibm-granite	●	22.4	0.027	124	1× B200 9B	2026-08	try
Qwen3 235B A22B Thinking 2507 qwen	●	22.1	0.0676		2× B200 235B, 22B active	2025-07	try
GPT-4 Turbo openai	●	21.5	8.33	113		2024-04	try
DeepSeek V3 0324 deepseek	●	21.2	0.0985	133	5× B200 685B, 32B active	2025-03	try
Kimi K2 Thinking moonshotai	●	21	0.0985	118	7× B200 1026B, 32B active	2025-11	try
gpt-oss-20b openai	●	20.7	0.012	219	1× B200 21B, 4B active	2025-08	try
Mistral Medium 3.1 mistralai	●	20.5	0.933			2025-08	try
GPT-4.1 Mini openai	●	20.2	0.8	148		2025-04	try
Mistral Large 3 2512 mistralai	●	20.1	0.833			2025-12	try
Llama 4 Maverick meta-llama	●	16.3	0.0522	86	3× B200 402B, 17B active	2025-04	try
GPT-5 Mini openai	●	15.6	0.833	98		2025-08	try
Qwen3 32B qwen	●	15.3	0.101	27	1× B200 33B	2025-04	try
Nemotron 3 Nano 30B A3B nvidia	●	14.4	0.00921		1× B200 30B, 3B active	2025-12	try
Ministral 3 14B 2512 mistralai	●	14.4	0.0428		1× B200 14B	2025-12	try
Qwen3 14B qwen	●	13.8	0.0454	59	1× B200 15B	2025-04	try
Nemotron 3 Nano Omni nvidia	●	13.8	0.00921		1× B200 30B, 3B active	2026-04	try
GPT-4 openai	●	13.1	40			2023-05	try
Llama 3.3 70B Instruct meta-llama	●	11.9	0.217		1× B200 71B	2024-12	try
GPT-4o-mini openai	●	11.4	0.3	81		2024-07	try
GPT-4.1 Nano openai	●	11.1	0.2	273		2025-04	try
Gemma 3 27B google	●	10.1	0.0842		1× B200 27B	2025-03	try
Ministral 3 8B 2512 mistralai	●	9.7	0.0274		1× B200 9B	2025-12	try
Llama 4 Scout meta-llama	●	8.2	0.0522	115	1× B200 109B, 17B active	2025-04	try
LFM2.5-2.6B liquid	●	7.7	0.00828		1× B200 3B	2026-08	try
Gemma 3 12B google	●	5.8	0.0374		1× B200 12B	2025-03	try
Llama 3.1 8B Instruct meta-llama	●	5.4	0.0247		1× B200 8B	2024-07	try
Ministral 3 3B 2512 mistralai	●	4.8	0.0118		1× B200 4B	2025-12	try
Gemma 3 4B google	●	2.7	0.0132		1× B200 4B	2025-03	try

how open weights are priced

A card is written off over 3 years at 50% utilization, plus electricity at $0.2/kWh, times 1.3 for the box around it. The cheapest card (at most 8 per box) whose memory holds the fp8 weights wins. An output token costs card-hours ÷ (peak TFLOPS × 0.2 ÷ 2 × active parameters), an input token the same at 0.6. No margin, no idle box, no engineer: a floor, not a quote. Parameter counts come from Hugging Face; for mixture-of-experts models the active share is estimated.

card	$	GB	TFLOPS	W	$ / h
RTX 5090	2000	32	210	575	0.50
H100	30000	80	989	700	3.33
B200	40000	180	2250	1000	4.48