Lernex/Metis-1.6-Logos-Base

Technical report  ·  Metis-1.6  ·  Lernex  ·  X

Metis-1.6 Logos Base

Logos Base is the larger pretrained model of Metis-1.6: a 10.23B-parameter mixture of experts that interleaves Mamba-2 and attention, runs every token through its stack twice, and touches 1.18B parameters on each pass. I trained it from scratch on 804.76B tokens, then extended its context to 131,072 tokens with 17.94B more. It continues text; it has not been taught to follow instructions or chat. For that, use Metis-1.6 Logos, which ranks second of 30 open models in mathematics.

Benchmarks

The historical 4K Logos checkpoint beside Praxis and 14 base models from other labs, every one run on this server through the same harness, prompts and scorers. The table preserves scores from the original 4K Logos run; the Context row shows the current 18B release window. New 18B checkpoint results are listed above. Praxis keeps the second column; the rest are ordered by Average, highest first. The labs' models are Logos's size: four mixtures of experts of 6 to 8.5B parameters that run 0.6 to 1.4B of them per token, seven dense models of 7 to 12B, and three 4B models from Praxis's table, between the two in compute per token. Logos runs 2.37B per token across its two passes. The picker above the table narrows it to the models you choose; bold is always the best score among the models in view. … is queued and - does not apply. Updated October 12, 2026, 18:50 UTC; the table fills in as runs finish.

Metis-1.6 Logos Basevsall 15 modelsPraxis BaseQwen3 8BQwen3.5 9BMinistral 3 8BFalcon-H1 7BQwen3.5 4BQwen3 4BGranite 4.1 8BGemma 4 12BGranite 4.0 H TinyGemma 4 E4BOlmo 3 7BTrinity NanoLFM2.5 8B-A1BOLMoE 1B-7B

Pick any to compare with Logos; pick none to see every model. Praxis first, then best first by Average.

Praxis Base2.94B · avg 24.66Qwen3 8B8.19B · avg 44.65Qwen3.5 9B8.95B · avg 44.41Ministral 3 8B8.49B · avg 43.08Falcon-H1 7B7.59B · avg 43.00Qwen3.5 4B4.21B · avg 42.93Qwen3 4B4.02B · avg 42.88Granite 4.1 8B8.38B · avg 41.54Gemma 4 12B11.94B · avg 41.19Granite 4.0 H Tiny6.94B · avg 39.01Gemma 4 E4B7.52B · avg 38.55Olmo 3 7B7.30B · avg 38.16Trinity Nano6.12B · avg 36.48LFM2.5 8B-A1B8.47B · avg 35.46OLMoE 1B-7B6.92B · avg 29.44

	Metis-1.6 Logos Base	Metis-1.6 Praxis Base	Qwen3 8B	Qwen3.5 9B	Ministral 3 8B	Falcon-H1 7B	Qwen3.5 4B	Qwen3 4B	Granite 4.1 8B	Gemma 4 12B	Granite 4.0 H Tiny	Gemma 4 E4B	Olmo 3 7B	Trinity Nano	LFM2.5 8B-A1B	OLMoE 1B-7B
# Params	10.23B	2.94B	8.19B	8.95B	8.49B	7.59B	4.21B	4.02B	8.38B	11.94B	6.94B	7.52B	7.30B	6.12B	8.47B	6.92B
# Activated params, per pass	1.18B	0.464B	6.95B	6.92B	7.42B	6.79B	3.57B	3.63B	7.97B	10.94B	1.31B	4.03B	6.48B	0.614B	1.42B	1.08B
# Activated params, total	2.37B	0.94B	6.95B	6.92B	7.42B	6.79B	3.57B	3.63B	7.97B	10.94B	1.31B	4.03B	6.48B	0.614B	1.42B	1.08B
Context	131K	128K	32K	256K	256K	256K	256K	32K	128K	256K	128K	128K	64K	128K	125K	4K
Training tokens	0.82T	0.82T	36T	undisclosed	1-3T	12T	undisclosed	36T	14.9T	undisclosed	22.5T	undisclosed	6.08T	10T	38T	5.13T
Q&A or instruction data in pretraining	none	none	yes	undisclosed	undisclosed	yes	undisclosed	yes	yes	undisclosed	undisclosed	undisclosed	yes	yes	undisclosed	yes
Obeys a bare instruction	0.0%	5.5%	90.5%	80.5%	0.5%	3.5%	36.0%	81.5%	23.5%	7.0%	53.0%	4.0%	9.5%	0.0%	2.5%	0.0%
Overall
Average	31.03	24.66	44.65	44.41	43.08	43.00	42.93	42.88	41.54	41.19	39.01	38.55	38.16	36.48	35.46	29.44
Average, with long context	34.99	27.72	-	46.05	46.75	46.55	45.57	-	45.20	45.11	42.10	41.54	-	39.64	-	-
Frontier reasoning and knowledge
Humanity's Last Exam	4.59	3.89	5.61	2.87	5.33	5.24	3.34	3.99	5.19	4.77	4.17	4.22	6.67	5.42	4.54	4.73
HLE-Diamond	3.71	3.43	4.40	2.88	4.26	4.26	2.88	5.08	6.18	5.49	4.12	3.71	5.91	5.22	3.71	5.22
HLE-Verified (Gold)	5.91	4.00	5.74	3.65	5.91	6.61	4.00	4.35	6.09	6.43	3.83	5.04	7.65	6.26	5.04	5.57
GPQA	30.36	27.90	39.96	38.84	37.95	37.05	25.67	22.99	34.60	35.94	31.25	26.12	30.36	31.70	26.12	26.79
GPQA Diamond	27.27	24.24	36.36	33.84	37.37	38.89	24.24	25.25	35.86	36.36	30.81	27.27	32.32	31.31	22.73	23.23
SuperGPQA	15.29	11.12	33.21	34.49	30.63	27.68	27.11	28.07	23.28	26.99	19.16	24.71	21.65	21.36	23.53	13.55
MMLU-Pro	22.45	14.19	48.89	54.68	50.18	49.10	48.62	45.74	38.66	44.84	33.11	38.19	33.87	34.31	39.15	18.53
BIG-Bench Extra Hard	4.60	1.66	10.07	3.27	8.14	5.31	2.17	7.92	4.91	5.31	5.69	2.19	7.65	6.22	1.17	7.92
ARC-AGI-2	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
SimpleQA Verified	3.40	2.70	4.40	6.70	5.20	4.10	4.80	3.20	4.50	5.80	4.50	4.00	4.00	4.80	5.30	3.50
Math
AIME 2026	0.00	0.00	3.33	0.00	0.00	3.33	0.00	3.33	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00
HMMT Feb 2026	0.00	0.00	0.00	6.06	0.00	0.00	0.00	3.03	0.00	0.00	3.03	0.00	0.00	0.00	0.00	0.00
AIME 2025	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	0.00	3.33	0.00	3.33	0.00	0.00	0.00
MATH-500	26.60	16.20	52.20	52.20	42.40	53.20	44.00	48.20	44.60	31.20	40.80	20.60	39.80	36.40	41.00	14.40
Omni-MATH	6.93	5.83	13.46	12.71	10.64	13.48	14.79	13.30	11.16	7.16	10.19	6.14	10.14	9.01	10.46	4.56
GSM8K	40.03	22.59	83.17	86.20	75.06	48.29	80.89	72.33	75.36	62.24	70.96	56.41	71.72	55.72	67.02	50.64
GSM8K-Platinum	42.10	23.57	85.61	88.67	77.58	48.97	81.97	75.52	76.92	64.68	74.36	57.24	73.37	58.06	71.63	52.52
MGSM	18.76	6.25	65.71	68.84	69.20	59.64	50.98	51.13	56.29	63.42	48.40	43.20	39.09	34.84	44.04	14.22
Coding
LiveCodeBench	7.31	0.00	10.97	20.40	0.63	15.94	16.59	15.44	7.26	9.14	6.17	13.68	10.40	7.77	6.40	0.46
HumanEval+	29.88	16.46	48.78	42.68	25.00	53.05	44.51	51.22	46.95	30.49	35.98	25.00	34.76	24.39	43.29	7.93
MBPP+	30.69	35.98	61.11	57.41	54.23	58.99	72.49	80.42	59.52	49.74	60.05	65.34	50.26	50.53	48.15	24.34
CRUXEval-O	23.39	12.49	40.38	41.67	35.97	41.22	30.32	28.66	39.73	35.69	34.26	24.29	30.70	29.18	24.00	17.59
CRUXEval-I	23.66	20.84	44.84	45.14	38.11	43.28	48.96	44.00	42.58	40.23	31.04	37.13	31.31	30.32	18.06	16.20
BigCodeBench	13.42	7.28	25.35	23.68	15.26	23.16	35.70	41.84	0.44	21.49	0.44	31.84	16.67	14.39	11.93	5.61
MultiPL-E	18.25	8.80	41.83	39.50	32.50	34.75	32.02	41.82	37.08	4.00	39.17	18.62	23.75	24.17	3.50	9.67
SciCode	0.69	0.00	0.69	1.03	1.03	1.03	0.00	0.00	0.69	0.34	0.69	0.00	0.69	0.69	0.69	0.69
Long context
RULER (4K)	84.48	75.54	83.35	60.80	93.28	90.31	87.17	92.41	88.48	89.55	86.68	88.99	90.66	83.01	65.06	66.87
RULER (8K)	80.74	66.86	82.07	60.39	91.73	91.95	62.09	75.79	87.09	88.09	80.47	76.74	82.90	77.14	58.35	-
RULER (16K)	74.57	61.53	74.97	56.22	91.57	86.25	65.65	70.17	83.08	85.03	77.07	74.40	75.61	70.42	45.69	-
RULER (32K)	73.25	56.23	66.10	52.31	80.74	77.32	59.92	61.31	78.04	82.68	71.89	70.90	70.95	66.38	28.82	-
RULER (64K)	70.50	50.76	-	50.85	75.56	75.77	89.25	-	77.31	80.66	65.41	70.38	65.04	71.92	12.91	-
RULER (128K)	60.14	42.38	-	91.92	71.79	72.72	80.58	-	74.94	78.88	68.27	60.99	-	62.88	-	-
LongBench v2	25.05	23.26	32.21	36.58	30.62	31.41	32.01	31.61	32.01	27.83	25.25	27.24	31.01	28.23	25.84	26.44
Knowledge and languages
TriviaQA	60.17	38.29	64.94	65.07	70.36	63.85	50.58	50.62	61.47	76.06	66.51	64.93	59.77	70.11	57.24	65.50
MMMLU	31.93	28.57	58.34	62.12	58.27	55.07	59.91	57.26	50.16	58.70	49.46	56.91	37.89	39.77	47.16	30.43
TruthfulQA MC2	37.65	37.49	41.46	42.19	38.24	37.61	52.30	53.38	39.14	35.38	42.82	44.98	38.69	37.21	45.62	35.98
AGIEval (English)	30.31	18.46	54.85	60.15	48.18	50.58	44.83	45.92	41.47	46.81	39.33	36.45	35.96	31.38	27.02	28.22
General
MMLU	52.53	39.62	74.59	77.35	72.31	73.44	74.21	71.19	68.18	70.57	64.87	66.54	62.18	63.79	61.86	53.55
MMLU (5-shot)	54.80	43.26	76.70	71.91	76.67	77.46	72.97	73.07	71.73	73.89	66.19	70.65	64.64	66.07	63.27	55.11
MMLU (cloze)	42.71	38.80	47.99	49.18	51.71	49.74	48.41	49.32	47.22	51.98	46.77	48.72	45.41	46.12	48.05	41.76
MMLU-Redux	54.77	41.24	77.30	79.39	75.09	75.68	78.48	75.85	70.14	72.05	67.49	69.83	63.91	65.82	65.46	55.42
MMLU-Redux (5-shot)	56.11	44.41	78.37	72.44	78.14	78.49	76.44	77.47	73.07	74.96	68.68	73.38	66.63	67.98	67.43	56.61
BIG-Bench Hard	36.23	31.18	75.72	81.03	77.16	77.71	80.45	71.37	74.06	74.72	59.76	62.74	63.75	48.90	45.14	33.76
DROP	41.28	29.91	58.96	53.67	69.17	65.80	67.88	60.14	71.68	65.02	62.90	56.80	66.46	58.66	44.78	50.87
HellaSwag	74.69	66.53	79.35	80.35	79.33	78.58	75.23	73.75	80.96	81.92	78.59	78.51	75.01	75.88	73.81	78.00
PIQA	79.43	76.33	79.98	80.58	80.79	81.72	79.22	78.07	80.74	82.64	80.74	80.74	79.11	81.94	77.80	79.92
ARC-Easy	74.66	65.78	79.50	79.08	86.45	82.37	76.22	76.18	81.69	86.83	82.20	85.19	81.10	80.51	82.28	77.40
ARC-Challenge	45.22	38.23	55.89	56.74	61.69	59.90	55.29	51.96	54.78	62.12	54.10	59.73	53.33	56.06	59.47	49.49
WinoGrande	66.54	61.64	68.43	69.69	70.40	71.19	71.03	70.96	69.22	70.64	61.25	70.24	66.22	66.14	65.27	65.98
OpenBookQA	42.60	40.40	42.80	45.00	45.80	45.80	42.40	41.00	45.60	46.80	42.80	46.00	42.20	43.80	42.20	42.60

Reading the table

Columns. Every column is a base model, never an instruct one, run on this server through the same harness, prompts, shot counts, windows and scorers as Logos, in BF16 with greedy decoding wherever Logos decodes greedily. Logos stays in the first column, boxed, and Praxis, the smaller Metis-1.6 trained on the same data and tokens, in the second, its header in its own card's blue; the rest are ordered by their Average, highest first, and a column still running joins the order when its Average lands. Picking models in the menu above the table narrows it to Logos and those models, and the bold moves to the best of what is in view; unpicking them all brings every column back. Each header links to that model's weights. … is queued; - does not apply, such as a RULER length past a model's context window or past what fits the 16 GB evaluation GPU.

Average. The unweighted mean over the 44 benchmarks every finished column has a score for, so each average covers the same tests. The long-context tests are left out: RULER past 4K, since not every model can run that long, and LongBench v2, whose documents are cut to each model's own window, so its score measures how far a model can read as much as what it knows.

Average, with long context. The same 44 plus LongBench v2 and RULER at 8K, 16K, 32K, 64K and 128K, for the models that can run at every length.

# Params. The language model alone. Vision and audio encoders and multi-token-prediction heads are not counted.

# Activated params. The weights one token runs through, counted one way for every column: token embeddings, the readout and per-layer embedding tables are lookups rather than compute and are left out, and a mixture of experts counts only the experts a token is routed to. Per pass is one trip through the stack; total is a token's whole path. Logos and Praxis make two passes, each through every non-expert weight plus 4 routed experts per layer, of 192 in Logos and 128 in Praxis. LFM2.5 routes each token to 4 of 32 experts, Granite 4.0 H Tiny to 6 of 64 beside a shared expert, Trinity Nano to 8 of 128 beside a shared expert, and OLMoE to 8 of 64; every other model is dense and makes one pass. Labs that quote active parameters usually count the embeddings in, so their figures run higher: Trinity Nano's 1B is 0.61B here plus 0.41B of embedding and readout.

Training tokens. As each lab publishes it for the base model, its stages summed where it gives them. Logos: 804.76B of pretraining. Praxis: the same 804.76B, then 18B of context extension. Granite 4.0 H Tiny's four stages sum to 22.5T (its card rounds to 23T), Granite 4.1's five to 14.9T and Olmo 3's three to 6.08T; OLMoE's 5.13T includes its 100B-token anneal. LFM2.5, Trinity Nano and Falcon-H1 do not say whether their context-extension stages are inside the figure they give. Qwen3's 36T is the figure its report gives for every size and Ministral 3's 1-3T the range its paper gives for the family; Qwen3.5 and Gemma 4 publish no figure. Ministral 3 was distilled from a larger model's output distributions, which carry more per token than text alone.

Q&A or instruction data in pretraining. What each lab says went into its base model's pretraining, in its own report or card. Metis-1.6 Logos and Praxis: no question-answer, instruction or chat data; 87.9B of their 804.76B pretraining tokens are synthetic teaching prose. OLMoE: its final 100B-token anneal is 16.7% FLAN instruction data, with question-answer pairs besides (paper). Olmo 3: its 100B-token mid-training mixes in Tulu 3 SFT and FLAN instruction data, synthetic question answering and thinking traces, to lay groundwork for post-training (report). Granite 4.1: chain-of-thought and synthetic instruction data, about a quarter of its third phase (blog). Qwen3: trillions of synthesized tokens including question answering and instructions (report). Falcon-H1: question-answer pairs, chain-of-thought solutions and synthetic QA (report). Trinity Nano: synthetic data made by turning web pages into question-answer pairs, and FLAN instruction data in its context extension (report). LFM2.5, Granite 4.0, Qwen3.5, Gemma 4 and Ministral 3 do not say; LFM2.5's mid-training targets reasoning, math and tool use, and Qwen3.5's card notes its chat-template tokens were trained in the base model. A base model that has seen these formats starts closer to an instruct model, which helps most on zero- and few-shot question answering.

Obeys a bare instruction. The same 200 lines for every column: one-line instructions whose answer takes no knowledge or skill (say a word, copy a sentence, change its case, answer yes or no), sent as the instruction and a newline, with no chat template and no examples. The figure is the share answered exactly on the first line, greedy. A model trained on documents alone tends to carry on the page instead, repeating the instruction or starting a post; one that has trained on instructions answers them.

Not on this card. 13 benchmarks that need instruction following, a judge model, tools or an agent loop join the table with the post-trained Metis-1.6. These comprehensive peer results are from the original Logos checkpoint trained on 4,096-token sequences, before context extension. Its RULER entries past 4K are marked -, and LongBench v2 reads each document cut to that original run window. The current 18B release supports 131K; its measured context results appear above.

How each benchmark is run

Logos's operating point. Native BF16, forced depth 2, four routed experts per token, deterministic routing, depth-memory gates at 1.0, no n-gram memory and its 4,096-token training window: the setting Praxis is scored at, at Logos's window. These are scores for that fixed setting, not for every depth or routing policy the architecture allows.

Praxis's column is the shipped Praxis Base at 131,072 tokens, read from the same result files as its own card.

Every other column runs in BF16 through vLLM, fed the harness's exact token ids. A model too large for the 16 GB evaluation GPU keeps part of its weights in host memory, which changes how fast it runs, not what it scores.

Harness. lm-evaluation-harness 0.4.13, except where a benchmark ships its own grader (LiveCodeBench, BigCodeBench, MultiPL-E, SciCode, BIG-Bench Extra Hard, ARC-AGI-2), which is used unchanged. Each benchmark runs at its standard shot count for base models. The exceptions are deliberate: TriviaQA at 5 shots, SuperGPQA and SimpleQA Verified in base-model form, and the competition math rows on MATH-500's 4-shot prompt so they compare with it directly. benchmark_suite.json holds every row with its protocol and every column's score.

Multiple choice, 0-shot, scored by answer likelihood. Length-normalized for HellaSwag, ARC, OpenBookQA and PIQA. MMMLU questions too long for the window count wrong.
GPQA, GPQA Diamond, MMMLU, TruthfulQA MC2, AGIEval (English), MMLU, MMLU-Redux, HellaSwag, PIQA, ARC-Easy, ARC-Challenge, WinoGrande and OpenBookQA

Multiple choice, 5-shot, scored by answer likelihood. SuperGPQA uses five fixed demonstrations in place of its generative official protocol, and counts the questions too long for the window wrong.
SuperGPQA, MMLU-Pro, MMLU (5-shot) and MMLU-Redux (5-shot)

Multiple choice with no options shown. Each answer's own text is scored after the question, 0-shot, normalized by its length in bytes. The other MMLU rows show lettered options and score the letter, which also tests whether a small model can tie a letter to its option; this one does not.
MMLU (cloze)

Text only, no judge model. Multiple choice scored by answer likelihood; short answers generated greedily and matched exactly or as an equivalent. Items too long for the window count wrong. Not comparable with official HLE scores, which are judged and include images.
Humanity's Last Exam, HLE-Diamond and HLE-Verified (Gold)

Greedy generation, answer matched. GSM8K and GSM8K-Platinum 5-shot, strict match. SimpleQA Verified 5-shot, normalized match, no grader model. TriviaQA 5-shot. MGSM 8-shot native chain of thought in all 11 languages, items that outgrow the window counted wrong. BIG-Bench Hard 3-shot chain of thought. DROP 3-shot F1.
SimpleQA Verified, GSM8K, GSM8K-Platinum, MGSM, TriviaQA, BIG-Bench Hard and DROP

Minerva 4-shot prompt, greedy. MATH-500 by exact match with a 256-token cap; AIME, HMMT and Omni-MATH checked by math_verify, Omni-MATH without its judge model.
AIME 2026, HMMT Feb 2026, AIME 2025, MATH-500 and Omni-MATH

Code run against the tests, greedy. HumanEval+ 0-shot, MBPP+ 3-shot, and BigCodeBench's complete split, all 1,140 tasks with 4 GB of memory each (the official 30 GB is sized for a larger machine; 99.5% of reference solutions pass here). Prompts are right-stripped, as the reference harnesses do.
HumanEval+, MBPP+ and BigCodeBench

Code, pass@1 from 10 samples at temperature 0.2. LiveCodeBench with its official base-model prompt on the January to April 2025 problems, CRUXEval 2-shot, MultiPL-E as the mean of 18 languages.
LiveCodeBench, CRUXEval-O, CRUXEval-I and MultiPL-E

The benchmark's own prompt and scorer. BIG-Bench Extra Hard 0-shot, items too long for the window counted wrong. ARC-AGI-2 as text grids, two attempts. SciCode without background, the reply starting at the step's function header.
BIG-Bench Extra Hard, ARC-AGI-2 and SciCode

RULER, 13 tasks at each length. Haystacks built with each model's own tokenizer: 500 items per task up to 16K, the first 200 at 32K, the first 100 at 64K and 128K. No model is run past its context window, or past what fits the 16 GB evaluation GPU at that length.
RULER

LongBench v2, 0-shot. All 503 questions in the official prompt, scored by the answer letter's likelihood. A prompt longer than the window is cut in the middle, as the official protocol does. Chance is 25.
LongBench v2

Logos, row by row

Contamination screening

Logos was trained on Praxis's data, which was screened against a pinned registry of 63 benchmark sources (203 evaluation splits). A document matching any held-out item exactly, after normalization, or on long shared n-grams (13 tokens for text, 12 for code) was removed whole. That covers most of the suite above, including LiveCodeBench through its sixth release, which spans the January to April 2025 window scored here. AIME 2026 and HMMT February 2026 were written after the training data.

Screening lowers the risk; it does not prove absence. Paraphrased or translated copies can pass n-gram filters, and benchmarks outside the registry (SuperGPQA, BIG-Bench Extra Hard, AGIEval, PIQA, BigCodeBench, ARC-AGI-2) were not screened.

Against Praxis

Logos and Praxis are Metis-1.6 at two sizes: one architecture, one data mix and the same 804.76B pretraining tokens, Logos with 3.5 times the parameters and 2.5 times the compute per token. The table below preserves the original 4K Logos scores against Praxis at the same point of training and same 4,096-token window, on the seven zero-shot tasks the first Metis-1.6 card reported:

Benchmark	Metric	Logos	Praxis at 4K	Gain	Praxis Base
MMLU	acc	54.01	39.46	+14.55	39.62
HellaSwag	acc_norm	75.19	67.41	+7.79	66.53
ARC-Easy	acc_norm	74.41	69.65	+4.76	65.78
ARC-Challenge	acc_norm	47.27	39.85	+7.42	38.23
OpenBookQA	acc_norm	42.20	41.20	+1.00	40.40
PIQA	acc_norm	79.49	77.91	+1.58	76.33
WinoGrande	acc	65.98	61.80	+4.18	61.64
Average		62.65	56.75	+5.90	55.51

Praxis at 4K is Praxis's checkpoint before context extension, with its n-gram rows repaired; Praxis Base is the shipped model after 18B tokens of extension, the second column of the table above. Every score here is from this server.

Every benchmark, Logos against Praxis Base

Benchmark	Praxis Base	Logos	Change
Average of these 7	55.51	62.65	+7.15
MMLU	39.62	54.01	+14.39
HellaSwag	66.53	75.19	+8.66
PIQA	76.33	79.49	+3.16
ARC-Easy	65.78	74.41	+8.63
ARC-Challenge	38.23	47.27	+9.04
WinoGrande	61.64	65.98	+4.34
OpenBookQA	40.40	42.20	+1.80

Praxis Base had 18B more tokens than Logos, all of them context extension, which its card measures at under a point on these benchmarks on average. The table grows as Logos's runs land.

18B-token context extension

This release uses the final Logos context-extension checkpoint, at optimizer step 1,968. The stage plan targeted 18B tokens and ended at 17,937,499,304 tokens when its final training bundle stopped short. It extends the 4,096-token pretraining model to a 131,072-token inference context, with training sequences up to 163,840 tokens. The n-gram memory tensors are removed from the released weights, matching the original Logos base treatment.

The stage promotion score narrowly favored the 12B checkpoint, 4.0400 to 4.0397. This repository follows the explicit selection of the end-of-stage 18B checkpoint. On the held-out context gate, 18B has the lowest full-context tail NLL of the three measured checkpoints; the 12B checkpoint has slightly higher needle top-1 recall.

Context gate, 131,072 tokens	Original 4K base	12B checkpoint	This 18B checkpoint
Needle top-1 accuracy	0.195	0.688	0.669
Needle target log-probability	-5.839	-1.671	-1.743
Full-context tail NLL	5.689	1.508	1.495
Short-context tail NLL	1.711	1.709	1.695
Base-window NLL	1.550	1.538	1.525

The gate measurements use 384 held-out records. Lower NLL is better. They are context-extension gate results, not a broad benchmark of long-document understanding.

Short-task results

These seven 0-shot tasks were measured with the same harness and setup for the original 4K base and both context-extension checkpoints. The 18B scores used the full-context configuration. Scores are percentages over each task test set.

Benchmark	Metric	Original 4K base	12B checkpoint	This 18B checkpoint
MMLU	accuracy	54.01	52.69	52.53
HellaSwag	normalized accuracy	75.19	74.30	74.69
PIQA	normalized accuracy	79.49	80.03	79.43
ARC-Easy	normalized accuracy	74.41	74.45	74.66
ARC-Challenge	normalized accuracy	47.27	45.31	45.22
WinoGrande	accuracy	65.98	65.43	66.54
OpenBookQA	normalized accuracy	42.20	43.20	42.60
Average		62.65	62.20	62.24

The average is the unweighted mean of the seven tasks. The context-extension stage changes it by -0.41 points from the original 4K base for 18B.

Long-context retrieval

RULER mean accuracy across 13 retrieval tasks. The 32K and 128K runs used subsets; sample counts are shown for comparability.

Context	Samples per task	Total samples	This 18B checkpoint	12B checkpoint
4,096	100	1,300	84.48	83.61
32,768	50	650	73.25	71.48
131,072	25	325	60.14	58.82

These are retrieval-suite scores, not a general measure of long-document reasoning. The 131,072-token run covered 25 examples for each of the 13 RULER tasks. The 18B benchmark run used n-gram-free weights; a 376,804-token gate probe measured bits per byte of 0.733898 with memory versus 0.733894 with its gates zeroed, with a maximum mean gate of 0.003159.

Quickstart

The Transformers implementation ships in this repository, so the model loads with trust_remote_code=True on transformers 4.57 or newer. It needs about 21 GB of memory in BF16. It is a base model, so give it the beginning of a document, not a question.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Lernex/Metis-1.6-Logos-Base"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto", trust_remote_code=True)

prompt = "The three laws of thermodynamics state that"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=200, do_sample=True, temperature=0.7, top_p=0.95)
print(tokenizer.decode(output[0], skip_special_tokens=True))

This implementation computes exactly what the native model computes, and it is the same code as the post-trained model's: the two checkpoints share every tensor name and shape. The vLLM and SGLang plugins and the llama.cpp converter in the Metis-1.6 Logos card read the same configuration; they were validated on the post-trained weights.

Architecture

Parameters	10.23B; 1.18B per pass, two passes per token
Architecture	20 layers of Mamba-2 and attention, width 2,560; 192 routed experts, 4 per token
Context	131,072 tokens at inference; trained at up to 163,840 tokens; attention has no positional encoding
Training	822.69B tokens: 804.76B pretraining plus 17.94B context extension
Vocabulary	65,536 tokens
Weights	BF16 with FP32 routers and state-space parameters, one file for Transformers and the native implementation

Metis-1.6 does not stack more layers. It runs the same 20 twice.

• Two passes, one set of weights. Every token goes through 40 sequence-mixing and expert stages while only 20 are stored. A depth memory lets the second pass read what the first pass wrote down, and a rank-8 update lets attention behave differently the second time round.
• Mostly Mamba-2. Attention sits at layers 5, 10 and 15 only, and it uses no positional encoding at all: the Mamba-2 layers already know where each token is. So only 3 of 20 layers keep a KV cache, memory grows slowly with context, and the context stretched from 4,096 to 131,072 tokens without touching position handling.
• A latent mixture-of-experts. Every layer ends in 192 routed experts and one shared expert that work in a 1,024-dimensional latent space, four routed experts per token. That is where the efficiency comes from: most of the 10.23B parameters sit idle for any given token.
• Four residual streams of width 2,560, read, written and mixed at every sub-layer by a doubly stochastic 4x4 matrix that knows which pass it is on.
• A memory the model switched off. Pretraining included a 1.77B-parameter hashed n-gram memory feeding layers 2 and 5. The model learned to close its gates, so I removed it before release: it changes held-out loss by less than 0.001 nats.

A quick note

Run it, build on it, and try to break it. Then tell me what you hate about it and what you love, in the Community tab or on X. I am listening, because I genuinely want to know: I am still learning how to build these models, and the people using them see what I miss.

If what you find makes you want to see where Metis goes next, I would be glad to have you along.

Giulianno Vollmer, founder of Lernex

Files

The same model.safetensors loads in Transformers and in the native Metis-1.6 implementation; for the native one, build the model inside ngram_free.without_ngram_memory(), since the native code always builds an n-gram memory and this archive has none. Every score on this page was produced from these weights.

File	What it holds
model.safetensors	The weights, routed experts in global expert order
eval_native_config.json	The single-device configuration behind every score on this page
tokenizer.json	The tokenizer, 65,536 tokens, identical to Praxis's
NGRAM_CANONICAL_IDS.json, .uint16	The canonical ID sidecar the native forward takes as an input
ngram_free.py	Builds the model without its n-gram memory
NGRAM_MEMORY_REMOVED.json	What was removed, and from which file
conversion_receipt.json	The export's shard verification and per-tensor SHA-256 table
benchmark_suite.json	Context gates, current 18B short-task results, RULER and n-gram gate probe
benchmark_suite_4k.json	Historical broad benchmark suite for the original 4K checkpoint
config.json, configuration_metis16.py, modeling_metis16.py	Hugging Face AutoConfig / AutoModelForCausalLM (trust_remote_code=True)
tokenizer_config.json, special_tokens_map.json, generation_config.json	Tokenizer and generation settings for Transformers

The export arrived without a configuration. eval_native_config.json is Praxis's, with every value the two families' training configurations disagree on taken from Logos's (width, depth, attention layers, heads, expert count and size, state-space groups); all 699 exported tensors match it by name and shape, and the parameter count comes out exact.

Without its n-gram memory

Metis-1.6 was trained with a hashed n-gram memory, added to the residual stream through a learned gate at layers 2 and 5. In Logos it holds 1.77B parameters, and, as in Praxis, the model learned to keep the gate shut. Measured on 300,000 characters of prose:

Injection	Mean gate (0 shut, 1 open)	Update, relative to the stream
Layer 2, pass 1	0.0004	0.26%
Layer 2, pass 2	0.0031	0.21%
Layer 5, pass 1	0.0001	0.12%
Layer 5, pass 2	0.0021	0.15%

Removing it moves bits per byte from 0.73441 to 0.73444. Praxis's 4K checkpoint reads the same text at 0.7927.

So this archive has none of it: no tables, no projection, no gates, 1,774,549,800 parameters fewer. It is exactly the trained model with its n-gram gates at zero: the archive without the memory and the original with its gates at zero give bit-identical hidden states, bits per byte and greedy text. The original 4K scores in the broad comparison and Praxis sections were measured on this n-gram-free base. The context-extension tables above report the selected 18B checkpoint, also with its n-gram memory tensors removed.

Removed full-weight objects

The current 18B release contains no n-gram memory tensors. The original 4K model weights and the full-table 18B export were permanently removed from this repository history to reduce private storage. The active release weights are the 20.8 GB n-gram-free file; the full source export SHA and tensor-level provenance remain recorded in the removal manifest and conversion receipt.

Checkpoint identity

	SHA-256
Context-extension checkpoint	92db097ab8806993d4a26070adc37e904241010d5bdd77bdb6c68a281ed28ae5
Parent checkpoint	ebddf9827d1a6e29570a5460b332de3979cc287f95378c912c7e90253234e032
Stage configuration	501a745878846542e4a0ea1535880d3ab0d938b4f4c65d51f4cf757c972261b1
Source full checkpoint weights	df252a5131cb74a938102a67b9d3e800f116205e13759f4fb6738c04703929d6
Released model.safetensors	7db04339c72952493788ab793190de3f275a422b13870f2e63c8c4f8cdda837f

The source checkpoint contains 699 tensors and 12,167,777,697 serialized elements. The released weights omit 1,774,549,800 elements across 20 n-gram memory tensors, leaving 679 tensors and 10,393,227,897 serialized elements. This count includes the tied embedding and readout twice, plus buffers, and is larger than the parameter count. conversion_receipt.json records the source shard verification and per-tensor hashes; NGRAM_MEMORY_REMOVED.json records the removal and released file hash.

Citation

If Metis helps your work, please cite the technical report, and the model if you used it directly:

@misc{vollmer2026thinktwice,
  title         = {Think Twice: The Metis-1.6 Technical Report},
  author        = {Vollmer, Giulianno},
  year          = {2026},
  eprint        = {2610.37903},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2610.37903}
}

@misc{vollmer2026logosbase,
  title     = {{Metis-1.6 Logos Base}},
  author    = {Vollmer, Giulianno},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/Lernex/Metis-1.6-Logos-Base}
}

License

Metis-1.6 Logos Base is released under the Apache 2.0 license.