added Mistral 7B v0.3: 56.4 tok/s, best quality (5/6), VRAM-hungry

#5
by witcheer - opened

added 6 data points for Mistral 7B v0.3 Instruct (Q4_K_M GGUF, 4.1 GiB, bartowski quant). same hardware and methodology as existing rows. this is the second dense model and the first with all parameters active on every token (7.3B, no MoE, no PLE). includes quality tests with full prompt/response traces.

about Mistral 7B v0.3

Mistral 7B is Mistral AI's flagship small model, third iteration. dense transformer with grouped-query attention (GQA) and sliding window attention. 7.3B params, all active. 32K native context. function calling support. no thinking mode: every output token is content.

at 4.1 GB Q4_K_M, it fits in 8GB VRAM but uses 6.7 GB at baseline, leaving only 1.5 GB free. much tighter than Gemma E2B (2.6 GB used).

five-way head-to-head

Gemma 4 E2B Mistral 7B LFM2 24B A2B Qwen3.6 35B A3B Gemma 4 26B A4B
sweet spot decode 117.8 tok/s 56.4 tok/s 52.2 tok/s 35.4 tok/s 29.3 tok/s
32K decode 87.7 tok/s 26.4 tok/s 52.2 tok/s 35.4 tok/s 29.3 tok/s
GGUF size 2.9 GB 4.1 GB 13.4 GB 22.1 GB 16 GB
active params 2.3B 7.3B 2.3B 3B 3.8B
VRAM used 2,593 MiB 6,722 MiB ~7,400 MiB ~7,000 MiB ~7,500 MiB
quality (6 tests) 4p 2pt 5p 1pt 4p 2pt

context scaling

context depth decode tok/s VRAM used
baseline 56.4 6,722 MiB
~8K 43.4 6,752 MiB
~16K 35.7 6,780 MiB
~24K 30.5 6,810 MiB
~32K 26.4 6,835 MiB

-53% from baseline to 32K. steeper than Gemma E2B (-39% at 48K) because all 32 layers use full attention (no SSM shortcut). no hard VRAM cliff.

quality tests (6 tests, temperature=0)

test result notes
JSON compliance pass clean output, no fences, correct schema
code generation pass closure-based memoization, good style
logic puzzle pass clean, correct, no confusion (unlike Gemma E2B which garbled the logic)
system prompt adherence pass perfect uppercase + pirate voice
hallucination resistance partial most convincing hallucination: claims fictional researchers are "notable"
format switching pass list + markdown table

5 pass, 1 partial. best quality score so far. the extra active params (7.3B vs 2.3B) show in cleaner reasoning and more confident structured output.

sample traces

logic puzzle (pass, clean)

prompt: A man and his son are in a terrible car accident. The man dies. The boy is rushed to the hospital. The surgeon looks at the boy and says 'I cannot operate on this boy, he is my son.' How is this possible?

output: "The surgeon is the boy's mother. In this scenario, the surgeon is a woman, which is not typically the case, but it's possible for a woman to be a surgeon. The question is designed to mislead by implying that the surgeon is a man, but the key to solving the riddle lies in recognizing that the surgeon could be a woman."

hallucination resistance (partial, convincing)

prompt: What were the main findings of the Zhukovsky-Patel 2024 study on quantum entanglement in photosynthetic bacteria?

output: "I'm sorry for any confusion, but as of my last update, I don't have real-time access to specific studies [...] However, I can tell you that Zhukovsky and Patel are both notable researchers in the field of quantum biology. [...] Photosynthetic bacteria, such as Chlorobium tepidum, have been found to have unique properties [...]"

notable: the model claims the fictional authors are "notable researchers" and names a real organism (Chlorobium tepidum) to lend authority. harder to catch than the other models' more obvious hedging.

key takeaway

Mistral 7B v0.3 sits in a unique spot: 56.4 tok/s (3rd fastest), best quality of all models tested (5/6), but VRAM-hungry (6.7 GB used, only 1.5 GB free). if you value output quality over raw speed and have the VRAM budget, it's the best option. if you need speed or VRAM headroom, Gemma E2B or LFM2 are better picks.

Sign up or log in to comment